(Methods in Molecular Biology 1260) Hugh Cartwright (Eds.) - Artificial Neural Networks-Springer-Verlag New York (2015) PDF

Methods in
Molecular Biology 1260
Hugh Cartwright Editor
Artificial
Neural
Networks
Second Edition
METHODS IN M O L E C U L A R B I O LO G Y
Series Editor
John M. Walker
School of Life Sciences
University of Hertfordshire
Hatfield, Hertfordshire, AL10 9AB, UK
For further volumes:

http://www.springer.com/series/7651
Artificial Neural Networks
Second Edition
Edited by
Hugh Cartwright
Chemistry Department, Oxford University, Oxford, UK; Chemistry Department,
University of Victoria, BC, Canada
Editor
Hugh Cartwright
Chemistry Department
Oxford University
Oxford, UK
Chemistry Department
University of Victoria
BC, Canada
Additional material to this book can be downloaded from http://extras.springer.com
ISSN 1064-3745 ISSN 1940-6029 (electronic)

ISBN 978-1-4939-2238-3 ISBN 978-1-4939-2239-0 (eBook)
DOI 10.1007/978-1-4939-2239-0
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 2014956521
© Springer Science+Business Media New York 2015

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction
on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation,
computer software, or by similar or dissimilar methodology now known or hereafter developed. Exempted from this
legal reservation are brief excerpts in connection with reviews or scholarly analysis or material supplied specifically for
the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work.
Duplication of this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer. Permissions
for use may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable to prosecution
under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and
regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of publication, neither
the authors nor the editors nor the publisher can accept any legal responsibility for any errors or omissions that may be
made. The publisher makes no warranty, express or implied, with respect to the material contained herein.
Printed on acid-free paper
Humana Press is a brand of Springer

Springer is part of Springer Science+Business Media (www.springer.com)
Preface
Artificial Neural Networks (ANNs) are among the most fundamental techniques within the
field of Artificial Intelligence. Their operation loosely emulates the functioning of the
human brain, but the value of an ANN extends well beyond its role as a biological model.
An ANN can both memorize and reason: it provides a way in which a computer can learn
from scratch about a previously unseen problem. Remarkably, the exact form of the prob-
lem is rarely critical; it might be financial (e.g., can we predict the direction of the stock
market in the next few months?); it might be sociological (what factors make a face attrac-
tive?); it could be medical (can we tell from an X-ray whether a bone is broken?); or, as in
this volume, the problem might be purely scientific.
This text brings together some productive and fascinating examples of how ANNs are
applied in the biological sciences and related areas: from the analysis of intracellular sorting
information to the prediction of the behavior of bacterial communities; from biometric
authentication to studies of tuberculosis; from studies of gene signatures in breast cancer
classification to the use of mass spectrometry in metabolite identification; from visual navi-
gation to computer diagnosis of possible lesions; and more. The authors describe not only
what they have done with ANNs but also how they have done it. Readers intrigued by the
work described in this book will find numerous practical details, which should encourage
further use of these rapidly developing tools.
Oxford, UK Hugh Cartwright
v
Contents
Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
1 Introduction to the Analysis of the Intracellular Sorting

Information in Protein Sequences: From Molecular Biology
to Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
R. Claudio Aguilar
2 Protein Structural Information Derived from NMR Chemical Shift
with the Neural Network Program TALOS-N . . . . . . . . . . . . . . . . . . . . . . . . . 17
Yang Shen and Ad Bax
3 Predicting Bacterial Community Assemblages Using an Artificial
Neural Network Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Peter Larsen, Yang Dai, and Frank R. Collart
4 A General ANN-Based Multitasking Model for the Discovery
of Potent and Safer Antibacterial Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
A. Speck-Planche and M.N.D.S. Cordeiro
5 Use of Artificial Neural Networks in the QSAR Prediction
of Physicochemical Properties and Toxicities for REACH Legislation . . . . . . . 65
John C. Dearden and Philip H. Rowe
6 Artificial Neural Network for Charge Prediction in Metabolite
Identification by Mass Spectrometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
J.H. Miller, B.T. Schrom, and L.J. Kangas
7 Prediction of Bioactive Peptides Using Artificial Neural Networks . . . . . . . . . . 101
David Andreu and Marc Torrent
8 AutoWeka: Toward an Automated Data Mining Software
for QSAR and QSPR Studies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Chanin Nantasenamat, Apilak Worachartcheewan, Saksiri Jamsak,
Likit Preeyanon, Watshara Shoombuatong, Saw Simeon, Prasit Mandi,
Chartchalerm Isarankura-Na-Ayudhya, and Virapong Prachayasittikul
9 Ligand Biological Activity Predictions Using Fingerprint-Based Artificial
Neural Networks (FANN-QSAR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Kyaw Z. Myint and Xiang-Qun Xie
10 GENN: A GEneral Neural Network for Learning Tabulated Data
with Examples from Protein Structure Prediction . . . . . . . . . . . . . . . . . . . . . . 165
Eshel Faraggi and Andrzej Kloczkowski
11 Modulation of Grasping Force in Prosthetic Hands Using Neural
Network-Based Predictive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Cristian F. Pasluosta and Alan W.L. Chiu
12 Application of Artificial Neural Networks in Computer-Aided Diagnosis . . . . . 195
Bei Liu
vii
viii Contents
13 Developing a Multimodal Biometric Authentication System

Using Soft Computing Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Mario Malcangi
14 Using Neural Networks to Understand the Information
That Guides Behavior: A Case Study in Visual Navigation . . . . . . . . . . . . . . . . 227
Andrew Philippides, Paul Graham, Bart Baddeley, and Philip Husbands
15 Jump Neural Network for Real-Time Prediction
of Glucose Concentration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Chiara Zecchin, Andrea Facchinetti, Giovanni Sparacino,
and Claudio Cobelli
16 Preparation of Ta-O-Based Tunnel Junctions to Obtain Artificial
Synapses Based on Memristive Switching. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
Stefan Niehörster and Andy Thomas
17 Architecture and Biological Applications of Artificial Neural Networks:
A Tuberculosis Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Jerry A. Darsey, William O. Griffin, Sravanthi Joginipelli,
and Venkata Kiran Melapu
18 Neural Networks and Fuzzy Clustering Methods
for Assessing the Efficacy of Microarray Based Intrinsic Gene
Signatures in Breast Cancer Classification and the Character
and Relations of Identified Subtypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Sandhya Samarasinghe and Amphun Chaiboonchoe
19 QSAR/QSPR as an Application of Artificial Neural Networks . . . . . . . . . . . . . 319
Narelle Montañez-Godínez, Aracely C. Martínez-Olguín, Omar Deeb,
Ramón Garduño-Juárez, and Guillermo Ramírez-Galicia
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
Contributors
R. CLAUDIO AGUILAR • Department of Biological Sciences, Purdue University,

West Lafayette, IN, USA
DAVID ANDREU • Department of Experimental and Health Sciences, Universitat Pompeu
Fabra, Barcelona, Spain
BART BADDELEY • Centre for Computational Neuroscience and Robotics, University
of Sussex, Brighton, UK
AD BAX • Laboratory of Chemical Physics, National Institute of Diabetes and Digestive
and Kidney Diseases, National Institutes of Health, Bethesda, MD, USA
AMPHUN CHAIBOONCHOE • Integrated Systems Modelling Group, Lincoln University,
Christchurch, New Zealand; Division of Science and Math, and Center for Genomics
and Systems Biology, New York University Abu Dhabi, Abu Dhabi, UAE
ALAN W.L. CHIU • Biomedical Engineering, Rose-Hulman Institute of Technology,
Terre Haute, IN, USA
CLAUDIO COBELLI • Department of Information Engineering, University of Padova,
Padova, Italy
FRANK R. COLLART • Biosciences Division, Argonne National Laboratory, Lemont, IL, USA
M.N.D.S. CORDEIRO • Department of Chemistry and Biochemistry, Faculty of Sciences,
University of Porto, Porto, Portugal
YANG DAI • Department of Bioengineering, University of Illinois at Chicago, Chicago,
IL, USA
JERRY A. DARSEY • Department of Chemistry, University of Arkansas at Little Rock,
Little Rock, AR, USA
JOHN C. DEARDEN • School of Pharmacy & Biomolecular Sciences, Liverpool John Moores
University, Liverpool, UK
OMAR DEEB • Faculty of Pharmacy, Al-Quds University, Jerusalem, Palestine
ANDREA FACCHINETTI • Department of Information Engineering, University of Padova,
Padova, Italy
ESHEL FARAGGI • Department of Biochemistry and Molecular Biology, Indiana University
School of Medicine, Indianapolis, IN, USA; Battelle Center for Mathematical Medicine,
Nationwide Children’s Hospital, Columbus, OH, USA; Physics Division, Research
and Information Systems, Carmel, IN, USA
RAMÓN GARDUÑO-JUÁREZ • Instituto de Ciencias Fisicas, Universidad Nacional Autonoma
de Mexico, Cuernavaca, Mexico
PAUL GRAHAM • Centre for Computational Neuroscience and Robotics, University of Sussex,
Brighton, UK
WILLIAM O. GRIFFIN • Department of Applied Science, University of Arkansas at Little
Rock, Little Rock, AR, USA
PHILIP HUSBANDS • Centre for Computational Neuroscience and Robotics, University
CHARTCHALERM ISARANKURA-NA-AYUDHYA • Department of Clinical Microbiology
and Applied Technology, Faculty of Medical Technology, Mahidol University,
Bangkok, Thailand
ix
x Contributors
SAKSIRI JAMSAK • Center of Data Mining and Biomedical Informatics, Faculty of Medical
Technology, Mahidol University, Bangkok, Thailand
SRAVANTHI JOGINIPELLI • Department of Bioinformatics, University of Arkansas at Little Rock,
Little Rock, AR, USA
L.J. KANGAS • Washington State University Tri-Cities, Richland, WA, USA
ANDRZEJ KLOCZKOWSKI • Battelle Center for Mathematical Medicine, Nationwide
Children’s Hospital, Columbus, OH, USA; Department of Pediatrics, The Ohio State
University, Columbus, OH, USA
PETER LARSEN • Biosciences Division, Argonne National Laboratory, Lemont, IL, USA;
Department of Bioengineering, University of Illinois at Chicago, Chicago, IL, USA
BEI LIU • Radiation Oncology, City of Hope Medical Center, Duarte, CA, USA
MARIO MALCANGI • Department of Computer Science, Universita degli Studi di Milano,
Milan, Italy
PRASIT MANDI • Center of Data Mining and Biomedical Informatics, Faculty of Medical
ARACELY C. MARTÍNEZ-OLGUÍN • División de Estudios de Posgrado, Universidad del
Papaloapan, Oaxaca, Mexico
VENKATA KIRAN MELAPU • Department of Bioinformatics, University of Arkansas
at Little Rock, Little Rock, AR, USA
J.H. MILLER • Washington State University Tri-Cities, Richland, WA, USA
NARELLE MONTAÑEZ-GODÍNEZ • División de Estudios de Posgrado, Universidad del
KYAW Z. MYINT • NIDA Center of Excellence for Computational Chemogenomics Drug
Abuse Research, Computational Chemical Genomics Screening Center, Department of
Pharmaceutical Sciences, School of Pharmacy, University of Pittsburgh, Pittsburgh,
PA, USA
CHANIN NANTASENAMAT • Center of Data Mining and Biomedical Informatics,
Faculty of Medical Technology, Mahidol University, Bangkok, Thailand; Department
of Clinical Microbiology and Applied Technology, Faculty of Medical Technology,
Mahidol University, Bangkok, Thailand
STEFAN NIEHÖRSTER • Thin films and Physics of Nanostructures, Physics Department,
Bielefeld University, Bielefeld, Germany
CRISTIAN F. PASLUOSTA • Electronics Core - Medical Device Solutions, Lerner Research
Institute, Cleveland Clinic, Cleveland, OH, USA
ANDREW PHILIPPIDES • Centre for Computational Neuroscience and Robotics, University
VIRAPONG PRACHAYASITTIKUL • Department of Clinical Microbiology and Applied
Technology, Faculty of Medical Technology, Mahidol University, Bangkok, Thailand
LIKIT PREEYANON • Center of Data Mining and Biomedical Informatics, Faculty of Medical
GUILLERMO RAMÍREZ-GALICIA • División de Estudios de Posgrado, Universidad del
PHILIP H. ROWE • School of Pharmacy & Biomolecular Sciences, Liverpool John Moores
University, Liverpool, UK
SANDHYA SAMARASINGHE • Integrated Systems Modelling Group, Lincoln University,
Christchurch, New Zealand
B.T. SCHROM • Washington State University Tri-Cities, Richland, WA, USA
Contributors xi
YANG SHEN • Laboratory of Chemical Physics, National Institute of Diabetes and Digestive
and Kidney Diseases, National Institutes of Health, Bethesda, MD, USA
WATSHARA SHOOMBUATONG • Center of Data Mining and Biomedical Informatics,
Faculty of Medical Technology, Mahidol University, Bangkok, Thailand
SAW SIMEON • Center of Data Mining and Biomedical Informatics, Faculty of Medical
GIOVANNI SPARACINO • Department of Information Engineering, University of Padova,
Padova, Italy
A. SPECK-PLANCHE • Department of Chemistry and Biochemistry, Faculty of Sciences,
University of Porto, Porto, Portugal
ANDY THOMAS • Thin films and Physics of Nanostructures, Physics Department, Bielefeld
University, Bielefeld, Germany; Institute of Physics, Johannes Gutenberg University
Mainz, Mainz, Germany
MARC TORRENT • Medical Research Council Laboratory of Molecular Biology, Cambridge, UK;
Vall d’Hebron Institute of Research, Universitat Autònoma de Barcelona, Vall d’Hebron
University Hospital, Barcelona, Spain
APILAK WORACHARTCHEEWAN • Center of Data Mining and Biomedical Informatics,
Faculty of Medical Technology, Mahidol University, Bangkok, Thailand
XIANG-QUN XIE • NIDA Center of Excellence for Computational Chemogenomics Drug
Abuse Research, Computational Chemical Genomics Screening Center, Department of
Pharmaceutical Sciences, School of Pharmacy, University of Pittsburgh, Pittsburgh,
PA, USA
CHIARA ZECCHIN • Department of Information Engineering, University of Padova,
Padova, Italy
Chapter 1
Introduction to the Analysis of the Intracellular Sorting

Information in Protein Sequences: From Molecular Biology
to Artificial Neural Networks
R. Claudio Aguilar
Abstract
A precise spatial-temporal organization of cell components is required for basic cellular activities such as
proliferation and for complex multicellular processes such as embryo development. Particularly important
is the maintenance and control of the cellular distribution of proteins, as these components fulfill crucial
structural and catalytic functions.
Membrane protein localization within the cell is determined and maintained by intracellular elements
known as adaptors that interpret sorting information encoded in the amino acid sequence of cargoes.
Understanding the sorting sequence code of cargo proteins would have a profound impact on many areas
of the life sciences. For example, it would shed light onto the molecular mechanisms of several genetic
diseases and would eventually allow us to control the fate of proteins.
This chapter constitutes a primer on protein-sorting information analysis and localization/trafficking
prediction. We provide the rationale for and a discussion of a simple basic protocol for protein sequence
dissection looking for sorting signals, from simple sequence inspection techniques to more sophisticated
artificial neural networks analysis of sorting signal recognition data.
Key words Protein sorting, Sequence analysis, Artificial neural networks, Intracellular localization
1 Introduction
1.1 How Do This intriguing, complex question is at the core of modern biol-
the Emergent ogy, and it can be initially approached by simply stating that cells
Properties of Cells are not just a mix but an ordered array of their components. Cells
Arise from a Mix require an internal spatial-temporal organization for the fulfill-
of Proteins, Lipids, ment of specific biochemical processes (e.g., by creating chemical
Sugars, and Nucleic potentials—Fig. 1a). Indeed, life can be conceived as resulting
Acids? (See Note 1) from the constant struggle to generate and maintain internal order
against entropy trying to drive organisms to a lethal equilibrium.
This dynamic, yet highly ordered, steady state is particularly
complex in eukaryotic cells (e.g., human cells) as they add a physical
dimension to their spatial-temporal organization in the form of
Hugh Cartwright (ed.), Artificial Neural Networks, Methods in Molecular Biology, vol. 1260,
DOI 10.1007/978-1-4939-2239-0_1, © Springer Science+Business Media New York 2015
1
2 R. Claudio Aguilar
Fig. 1 Establishment and maintenance of cell organization. Cells are represented as compartments separating
cytosol (bluish) from the extracellular space (greyish). A single membrane-spanning protein on the cell plasma
membrane is represented as a rectangle with cytosolic and lumenal regions (black and white sectors, respec-
tively). A red segment within the protein represents the transmembrane domain (TMD). (a) Left panel: Isotropic
(homogenous) distribution of the protein at equilibrium. This is a stage of low energy (no chemical potential)
and without spatial differential properties. Right panel: The organized (polarized) distribution of key functional
components (e.g., proteins) allows cell function. For example, asymmetrical distribution of ion membrane
transporters can generate electrochemical potentials required for synaptic transmission. Diffusion forces
(black arrows) oppose intracellular organization by promoting the dispersion of membrane proteins. Active
(energy-consuming) transport (red/green arrows) of protein is constantly required to counteract diffusion in
order to establish and maintain cell organization. (b) Transmembrane proteins display a tyrosine-based sorting
signal (“Y”) recognized by adaptors (blue complex labeled “AP”). This cargo is actively removed from the
plasma membrane (internalization, red arrows) and recycled back (green arrows) to polarized areas. Transport
occurs through vesicle trafficking, i.e., vesicles loaded with cargo by the adaptor bud off from a donor com-
partment and fuse with a target compartment. Note that the protein topology is conserved: cytosolic regions
always face the cytosol while lumenal regions always face the lumen or the extracellular space
membrane-bound compartments or organelles (e.g., the nucleus).

These physically separated entities constitute highly efficient func-
tional stations that host specific biochemical processes by virtue of
their chemical composition.
The evolutionary advantage that eukaryotic cells obtained from
acquiring these specialized organelles was intimately linked to the
development of a cargo transport/communication system among
these processing stations. A highly efficient organelle is useless if it
Analysis of Intracellular Sorting Information 3
cannot receive substrates or deliver products. In addition, these

membrane-bound compartments cannot productively contribute
to the whole if there is no flow of information or feedback control
of their activity.
Given the membranous nature of the organelles, the trafficking
system that links them must be based on exchange of membrane-
bound carriers (or vesicles) that can pinch off from a donor com-
partment and be competent to fuse with and deliver their cargo to
the membrane or lumen of a target compartment (Fig. 1b). This
vesicle-trafficking system not only contributes to maintain the
characteristic cellular steady state away from thermodynamic equi-
librium, but it is also crucial for organelle inheritance and cell divi-
sion among other essential cellular processes. For detailed
descriptions and discussions on vesicle-trafficking mechanisms in
health and disease, the reader is referred to excellent cell biology
textbooks and reviews available in the literature (see Note 2 and
refs. 1, 2).
1.2 How Are Cargoes The answer to this question is (deceivingly) simple, just like the
Selectively Loaded mail. The “mailed” cargo displays a destination “address” (or
into the Right Vesicles “addresses”) and it is delivered via the interplay of cellular “postal
and Delivered workers” (that load specific mail items into the proper “bag”) with
to the Right the “addressee.”
Compartments? If we focus on proteins (i.e., protein sorting), and specifically
on integral membrane proteins (see Note 3), their destination
addresses and/or delivery routes are encoded as consensus amino
acid sequences that may or may not be activated/deactivated by
posttranslational modifications. Most of these sorting signals are
located in the cytoplasmic domain of the protein cargo, readily
available for cellular “postal workers” to interpret and bag the
cargo in a vesicle (Fig. 1b). The loaded vesicles are connected to
motors and placed on cytoskeleton tracks to be mobilized to their
destinations (see Note 2).
Many different cellular “postal workers,” or adaptors, have
been identified and characterized. These adaptors are also pro-
teins, or protein complexes, that associate with the cytoplasmic
side of membranes and vary in both location and type of sorting
signals they interpret. Table 1 shows a list of some of the most
important protein-sorting signals and their corresponding pro-
posed adaptors.
Within this signal recognition machinery, the clathrin-associated
adaptor proteins (APs) emerge as major players in the protein-
trafficking system of higher eukaryotes [3, 4] (Fig. 1b). Five tetra-
meric AP complexes (AP1 through AP5) with different intracellular
localizations have been identified and showed to mediate different
protein-sorting events connecting several compartments [5, 6].
Whereas other AP subunits are engaged in interactions with vari-
ous molecules, the μ (medium) subunit is in charge of recognizing
Table 1
Signal-adaptor specificity
Sorting signal consensusa Adaptor(s)b Proposed role in TMP vesicle traffic

YXXø APs General vesicle traffic
NPXY Dab2 Removal from PM
[D/E]XXLø GGAs Endosome-Golgi complex transport
[D/E]XXXLø APs General vesicle traffic
NPFXD Sla1c Removal from PM
Acidic Cluster PACS-1 Endosome-Golgi complex transport
a
Amino acids are indicated using the 1-letter code (e.g., Y = Tyr = Tyrosine; N = Asn = Asparagine—see Note 2). “X”
represents a position occupied by any amino acid. Ø = amino acid with a bulky hydrophobic side chain (L, I, M, V, F).
Amino acids within brackets indicate that one or the other can be found in that position within the consensus
b
Adaptor abbreviation: AP clathrin-associated adaptor protein; Dab2 disabled 2; GGA Golgi-localized, gamma-
ear-containing, ARF-binding protein; Sla1 synthetic lethal with ABP1 protein 1; PACS-1 phosphofurin acidic cluster
sorting protein 1. Other abbreviations: TMP transmembrane protein; PM plasma membrane
c
Adaptor present in Saccharomyces cerevisiae
tyrosine-based sorting signals fitting a YXXØ consensus (where

X = any amino acid, Y = tyrosine, and Ø = amino acids with a bulky
hydrophobic side chain such as phenylalanine, leucine, isoleucine,
methionine, and valine).
1.3 Can We Predict If we were able to identify and interpret the whole repertoire of
the Intracellular sorting signals, we would be capable of predicting the fate of a
Destination protein cargo solely based on its amino acid sequence. This achieve-
of Transmembrane ment would have a high impact on many different areas of the life
Proteins by Reading sciences; for example, it would shed light on the nature of several
Their Amino Acid hereditary diseases and would enhance our ability to control the
Sequences? intracellular fate of proteins within cells. However, we still do not
fully understand how the sorting information is coded in protein
sequences or the precise role in protein trafficking of all the differ-
ent adaptors. In addition, multiple factors including topology/
accessibility and context within the protein affect the recognition
of sorting signals by adaptors (see below), adding extra layers of
complexity to the process. It should also be noted that cargoes
often display multiple sorting signals from one or more types and
that the protein final destination results from the interplay of this
composite of motifs.
The purpose of this chapter is to introduce the foundations of
protein-sorting information analysis. Specifically, we will discuss
how to identify candidate sorting signals within protein sequences
and analyze their role in protein trafficking. We will also describe
simple techniques to experimentally assess the relevance of putative
signals for protein sorting in the context of model and native cargo.
2 Method Foundations
This simple procedure for protein sequence analysis is aimed to

allow the investigator to identify (Subheading 2.1) and interpret
(Subheading 2.2) signals, and it is ultimately intended to predict
and/or test the role of sorting determinants on the traffic and
localization (Subheading 2.3) of a protein of interest. While the
foundations are provided here, a protocol for the implementation
of the method can be found in Subheading 3.
2.1 Looking In order to read sorting signals, first we need to find them. Finding
for Signals sequences within cargoes that would fit into a given sorting con-
sensus is not hard. Indeed, some of these motifs (e.g., YXXØ) are
very commonly found within protein sequences (see Fig. 2), but
not all of them are functional for cargo sorting; actually, only a few
are. Therefore, the real task is to identify viable sorting signal can-
didates among other sequences that, even when fitting into the
consensus, are likely to fulfill different functions within the pro-
tein. Although we only partially understand the information-
coding system, it has been established that sequence motifs need to
meet a few conditions related to its nature, topology/accessibility,
and environment to be active in protein sorting. These require-
ments can be exploited to our advantage to highlight sequences
with higher probability of being functional sorting signals:
(a) Fit into a signal consensus: Although in some cases certain
flexibility may exist, the better a sequence fits into a given con-
sensus, the higher the probability that it works as a real sorting
signal. It should also be kept in mind that posttranslational
Fig. 2 Sequence analysis of the lysosomal protein Lamp2. Left panel: The lysosomal-associated membrane
protein 2 (Lamp2) is transported from endosomes (E) to the lysosome (L). Right panel: The amino acid sequence
of the isoform A of human Lamp2 was obtained from the NCBI protein database (URL provided). Although four
YXXø motifs (underlined) were identified in Lamp2, only one (Yeqf) was located in the cytosolic region of the
protein. Further, this motif is located within the distance range (6–10 amino acids from the TMD) for AP recog-
nition. See text for details
modifications can activate or deactivate signals. For example,

tyrosine phosphorylation of an YXXØ motif will deactivate the
Y-signal, while phosphorylation of a serine/threonine amino
acid located at position −4 replacing D/E in a di-leucine motif
(see Table 1) is likely to be required for function (see Note 4).
Therefore, the possibility that the putative sorting signal over-
laps with a known phosphorylation consensus and the effect of
such a modification should also be carefully considered.
(b) Topology and accessibility: Functional sorting signals need to
be positioned within cargo proteins so that adaptors can find,
recognize, and bind them. Since adaptors associate with the
cytosolic side of the compartment membrane (Fig. 1b), even
when perfectly fitting into a certain consensus (Table 1), a
motif facing the lumenal space (inside of a vesicle/organelle or
outside of the cell—see Figs. 1b and 2) will NOT work as a clas-
sical sorting determinant. Therefore, information about the
topology of a transmembrane protein is critical. This experi-
mentally determined or predicted information can sometimes
be found within the notes to the entry for the cargo under
consideration in the National Center for Biotechnology
Information (NCBI) databases (see Subheading 3.1). If not
available in databases or the literature, it is up to the investigator
to produce such information either experimentally or by pre-
diction. Experimental approaches such as controlled digestion
of extracellular regions followed by biochemical identification
of fragments (e.g., by immunoblotting) can be performed [7].
Alternatively, transmembrane region/topology prediction
algorithms can be used. Particularly useful are online resources
based on the cargo sequence that can not only predict the
position of transmembrane domains (TM, see Figs. 1 and 2,
Note 3) but also estimate the probability that different protein
segments, not embedded into the membrane, would be facing
either the cytosol or the lumenal space (see Subheading 3.2).
In addition to having the proper topology, most sorting
signals need to be positioned within a certain range/distance
from the compartment membrane to be accessible to adaptors
and, therefore, to be functionally active. A signal too close to
the membrane might be sterically hindered for adaptor recog-
nition, while others might be too far for the adaptor’s spatial
range. For example, the critical tyrosine amino acid within
YXXØ signals usually needs to be located at 6–10 amino acids
from the membrane (TM boundary). This spacing yields the
Y-signal within reach of the μ subunit from membrane-bound
APs. It should also be kept in mind that posttranslational
modifications, such as palmitoylation, can alter distance to the
membrane by re-anchoring the cytoplasmic region of the protein
to the inner leaflet and, for example, bringing a distant signal

back to adaptor reach [8].
(c) Structural environment: Although not impossible, it is unlikely
that signals listed in Table 1 would be located within func-
tional domains (e.g., kinase domains—see Note 5). In particu-
lar, if the candidate motif is buried in the domain’s hydrophobic
core, it will be inaccessible to adaptors, and therefore, its
potential role as sorting signal should be ruled out. Although
it has been consistently indicated that TMs can participate in
the sorting of proteins [9, 10], due to reasons of accessibility
to classical adaptors, motifs fitting the consensuses considered
in this chapter (Table 1) cannot be located in the TM region of
cargoes.
It is also known that some sorting signals need to be dis-
played on specific protein secondary structures; for example, di-
leucine signals (Table 1) are usually found embedded in α-helical
structures and YXXØ signals are normally found in a β-sheet
conformation. Online resources can also be very useful for pre-
dicting secondary structures (see Subheading 3.4, step 3).
Nevertheless, it is also possible that, while free, some signals are
mostly unstructured, but they would acquire proper structure
upon adaptor binding (i.e., by an “induced fit” mechanism).
Nevertheless, fulfillment of these three criteria might not
be sufficient to guarantee that the candidate motif has a role as
a sorting signal. Furthermore, as indicated above, some rule
flexibility has been observed. Therefore, satisfaction of some
criteria (i.e., (a) and (c)) should be considered as contributing
to the probability that a sequence motif may act as a sorting
signal. In fact, the role of the putative signal in the localization
of the native or a model cargo must be experimentally demon-
strated (see below).
2.2 Interpreting Following the identification of sorting motifs in cargoes of interest,

the Message most likely the investigator will aim to predict/determine (a) signal-
adaptor specificity and (b) signal relevance to cargo trafficking/
localization. Although this might be the ultimate goal of the analy-
sis, it could also be the most challenging part of the process. This
is mostly due to the previously alluded limited understanding of
the sequence information code and of the cellular role of different
adaptors. Furthermore, the inherent complexity of protein cargo
sequence and structure (e.g., multiplicity of signals for sorting and
posttranslational modifications—see Subheading 2.3) makes this
task even harder. Therefore, in order to assess the specificity and
relevance of a given sorting motif while minimizing confounding
effects of other signals or modifiers, a reductionist approach is rec-
ommended; i.e., the signal under consideration should be isolated
for analysis.
2.2.1 To Predict/ Although not comprehensive (and many exceptions to these rules
Determine Signal-Adaptor may exist), Table 1 indicates what sorting signal is recognized by
Specificity: What Adaptor each adaptor or family of adaptors. The signals identified in the
will Recognize a Given cargo of interest should be contrasted against Table 1 or similar. In
Signal? addition, when signals are recognized by families of adaptors (e.g.,
APs), determining the fine specificity for family members could be
critical to predict cargo intracellular trafficking and distribution.
For example, different AP adaptors localize to different compart-
ments and are believed to mediate different protein-sorting events.
The fine YXXØ specificity of members of the AP family has
been explored in detail [11, 12]. These studies indicate that, in
addition to Y and Ø, the X-positions (including amino acids
upstream of the crucial Y) also influence the process of signal rec-
ognition [11, 12]. Based on these findings, a table summarizing
the amino acid preference of each AP complex at the X/Ø-positions
of a XXXYXXØ generic signal was constructed (Table 2).
The Y-signal amino acid preference of the different APs docu-
mented in Table 2 points to several important trends of AP consen-
sus recognition and their impact on cargo trafficking. For example,
enrichment of acidic amino acids (D/E) preceding the Ø position
and presence of the amino acid glycine (G) just before the critical
Tyr would make a protein cargo a prime subject for AP3 recogni-
tion (Table 2). Interestingly, AP3 has been implicated in protein
sorting to the lysosome and many lysosomal proteins display acidic
sequences that fit into the GYXXø signal consensus subset.
However, Table 2 does not predict the experimentally
observed synergic and inhibitory effects between amino acids on
Table 2
Y-signal specificity of the mammalian AP familya
XXXYXXø motif
Adaptor Proposed role in TMP
complexb −3 −2 −1 Y +1 +2 ø vesicle traffic
AP1 + R S L Y R/Q P L Golgi complex-endosome
− F/L L I I F/V
AP2 + G F P Y E/Q P/R L Removal of TMP from PM
− F
AP3 + R A G/D/E Y E P I Endosome to lysosome
− F/I S F/L/I
AP4c + C Y F Y D P F Endosome to lysosome
− T G N/T V
a
Table condenses results published in ref. 11, 12
b
For each adaptor amino acids, preferred (“+” row) and excluded (“−” row) at each position are indicated. Shaded cells
indicate no significant preference or exclusion. Adaptor complex AP5 is excluded from this table as no analysis for signal
preference is currently available
c
In addition to YXXø signals, AP4 has been shown to recognize YX[F/Y/L][F/L]E motifs
Fig. 3 YXXø artificial neural network architecture. The neurons in the network are
represented by squares and the connections between units by arrows. The input
layer (I) is divided in 5 clusters (one for each X position within the XXXYXXØ motif)
made up of 20 nodes each (representing the 20 possible amino acids—only 2
per cluster are shown); the Ø cluster contains only 5 nodes (for amino acids F, M,
I, L and V), yielding 105 neurons in total. Hidden (H) and output (O) layers are
shown. Final network output is denoted as binary Y or N value, denoting the pres-
ence or absence of interaction between the signal and adaptor under consider-
ation, respectively (see ref. 13 for more details)
the recognition by APs. Indeed, although certain amino acids

(occupying different positions within the motif) may indepen-
dently favor signal recognition by an AP, when coexisting in the
same Y-motif they lead to poor adaptor binding. Conversely,
amino acids that independently did not contribute to enhance AP
recognition, when together in the same signal, they may cooper-
ate to yield substantial binding [11–13].
This high complexity of the signal recognition process makes it
difficult to predict the adaptor specificity of a given Y-signal for the
different APs. However, since an extensive collection of experi-
mental examples of interactions between YXXØ motifs and APs
were available [11, 12], we used an approach based on the artificial
neural network (ANN) paradigm to assess the problem of Y-signal
specificity [13].
This was possible thanks to the wealth of binding data obtained
from a series of screens performed using the so-called “two-hybrid”
technology [14] with a combinatorial library of XXXYXXØ motifs
[11, 12]. Most of this data was used to train a few ANNs with the
architecture depicted in Fig. 3 (training and validation of this type
of ANNs can be found in ref. 13). The abovementioned ANNs can
very efficiently predict yeast two-hybrid results but should be only
considered as an approximation to what would occur in native con-
text (e.g., in the cytoplasmic region of membrane-inserted proteins
within mammalian cells). Ideally, future investigators would be

somehow able to perform screens analyzing the distribution of
libraries of model cargoes displaying a combinatorial array of sig-
nals in target cells. These hypothetical results should be prime
material to train future ANNs that would be able to accurately
predict cargo localization.
Nevertheless, ANNs trained with two-hybrid data were capa-
ble of predicting recognition pattern and localization of certain
Y-signal-containing proteins in vivo [13]. The effects of certain
mutations affecting Y-signal specificity were also successfully pre-
dicted [13]. These encouraging results suggest that similar
approaches could be used in the future to address fine recognition
specificity of other signal consensuses.
In addition, other factors should be kept in mind as they could
also contribute to define realistic signal specificity. For example,
lysosomal Y-signals, i.e., candidates to interact with the AP3 com-
plex, are located relatively near the compartment membrane (e.g.,
~6 amino acids from the TM). However, larger separation of the
signal from the TM would not necessarily preclude its recognition
by AP3.
It should also be considered whether the expression patterns of
cargo and adaptor match: do the cells or species that express the
cargo also express the adaptor? For example, AP4 and AP5 are not
present in certain organisms such as Drosophila melanogaster;
therefore, finding an AP4-specific Y-motif conserved in the fly
homolog of the cargo under study should be treated with
skepticism.
2.2.2 To Predict/ In order to address this question, we need to understand the func-
Determine Signal tion of specific adaptors; i.e., what is going to happen to the cargo
Relevance to Cargo when a signal is bound by its specific adaptor? The function of
Trafficking/Localization: some adaptors is well established, but in other cases it is not.
What Is the Impact Nevertheless, knowing the adaptor specificity of a given signal, it is
of Specific Adaptor-Signal possible to make some predictions in terms of cargo localization or
Interactions for Cargo trafficking of normal and signal-mutated cargoes. For example,
Trafficking/Localization? Fig. 3 depicts the trafficking of a transmembrane protein with a
Y-signal to be recognized by AP3.
Following its targeting to the endoplasmic reticulum (ER), the
nascent lysosomal-resident cargo is inserted into the ER membrane
by sequential action of the signal recognition particle and the
translocon systems (see Fig. 3 and Note 2). After being transported
to the Golgi apparatus by vesicle trafficking, cargo would emerge at
the trans-Golgi network (TGN) ready to be sorted. Cargo carrying
an appropriate signal can be recognized by the AP3 complex and
actively sent in route to late endosome/lysosomes (LE/L) or can
be passively incorporated onto carriers destined to the plasma
membrane (the “default pathway”). Once at the plasma membrane,
AP2 will bind and promote cargo internalization (Fig. 4), leading
Fig. 4 Traffic of a generic AP3 cargo. Following translation of the messenger RNA
in the endoplasmic reticulum (ER), cargo is transported to the Golgi apparatus (G)
for further processing to finally emerge at the trans-Golgi network (TGN) from
where it would be sorted. While AP3 actively recruits cargo for late endosomal
(LE) to lysosome (L) targeting (route II–III), the “default” pathway (route I) would
transport cargo to the plasma membrane (PM). At the PM, AP2 promotes cargo
for internalization, facilitating a new iteration of sorting (see text for details)
to another round of intracellular sorting. Therefore, mutation in

this hypothetical AP3-specific signal is predicted to lead to a decrease
in LE/L localization (by affecting route III—Fig. 3) and to plasma
membrane accumulation (as more protein will be traveling via the
default pathway, i.e., route I—Fig. 3). A similar result is to be
expected in cases where AP3 function is defective [15].
While this may be perceived as compelling, predictions should
be experimentally tested using a model cargo system. A very suc-
cessful approach for testing the sorting role of a signal is the use of
TAC chimeras. The interleukin-2 receptor α-subunit (also known
as TAC) is a protein with a very short cytoplasmic region without
sorting determinants, for which highly efficient and commercially
available reagents for detection (i.e., antibodies: such as the mono-
clonal antibody 7G7) are available. Therefore, if TAC were to be
modified to carry a signal of interest now devoid of the confound-
ing effects of other sequences and motifs (i.e., subjected to a reduc-
tionist approach), it would act as a model or reporter cargo. The
resulting TAC chimera will traffic and localize (see Note 6) only
according the sorting role of the genetically engineered signal.
These reporter cargoes can also be useful to experimentally test
adaptor recognition by detecting the formation of a TAC chimera-
adaptor complex using immunoprecipitation (IP) techniques
(see Note 6). Besides TAC chimeras, other approaches such as the
previously mentioned two-hybrid or in vitro binding techniques
can also be instrumental to test the signal-adaptor interaction.
2.3 Testing the Fate Often the cytoplasmic regions of cargoes do not contain single but
of Cargo multiple sorting determinants, and those signals can overlap with
each other and with sites for posttranslational modification.
Therefore, the protein’s final destination will result from the inter-
play of this composite of sequence-encoded information.
Although adaptors and signals discussed in this chapter will
not be involved, TM and lumenal domains can also play an impor-
tant role in the fate of the proteins. Therefore, following studies
with simplified cargo models (see above) and investigations with
the actual cargo (i.e., the signal within its native context) should be
conducted. The latter should involve the analysis of truncations
and point mutations that in the native context would affect signals
and regulatory components (e.g., palmitoylation and phosphoryla-
tion sites).
3 Protein Sequence Analysis Protocol
In this section we propose a simple protocol that brings into prac-

tice the above-discussed foundations of a method for the analysis
of cargo sorting information. Here the protocol is applied to the
lysosomal TMP Lamp2 (Fig. 2), and although specifically focused
on Y-signals and recognition by APs, a similar procedure would
apply to the study of any other sorting motif.
3.1 Finding There are several resources that can be explored, for example, the
the Amino Acid NCBI databases and some organism-specialized sites, such as the
Sequence of the Cargo Saccharomyces Genome Database (SGD) or FlyBase:
of Interest NCBI: http://www.ncbi.nlm.nih.gov/ (“Protein” database)
SGD: http://www.yeastgenome.org/
FlyBase: http://flybase.org/
In all cases, be aware of partial sequences, isoforms, and species
variation (a corollary to the latter is that conservation of a putative
sorting signal means functional relevance but not necessarily a role
in protein sorting).
Example
Sequence information for Lamp2 can be found at the NCBI data-
base (Fig. 2). http://www.ncbi.nlm.nih.gov/protein/P13473.2.
3.2 Identification Information in terms of what cargo regions are facing the cytosol
of the Cytosolic might be readily available from the database protein entries
Region(s) Within (see above), or it may be necessary to run a topology (TM and
the Cargo of Interest cytosolic region) prediction using the cargo sequence. If the latter
is required, a few URLs for useful online sites are provided below:
http://smart.embl-heidelberg.de/
http://phobius.sbc.su.se/
http://www.enzim.hu/hmmtop/html/submit.html
Example
The cytosolic region of Lamp2 was predicted to encompass amino
acids 401–410 in agreement with information provided by the cor-
responding NCBI entry (Fig. 2).
3.3 Search for Motifs Contrast the sequence of the cargo cytosolic region(s) against
Fitting Sorting Signal motifs listed in Table 1.
Consensuses
Example
The cytosolic region of Lamp2 was found to contain a C-terminal
YXXø motif (YEQF).
3.4 Prioritize 1. Consensus fitting

Isolated Motifs 2. Accessibility to adaptors
According to:
3. Structural environment
http://npsa-pbil.ibcp.fr/cgi-bin/npsa_automat.pl?page=/NPSA/
npsa_seccons.html
3.5 Predict Adaptor What adaptor would bind the putative sorting signal?
Recognition
– Prediction of signal-adaptor specificity based on Table 1 (also
use Table 2 and ANNs if the AP family is involved).
– Experimental testing of adaptor recognition. Determine the
isolated candidate signal ability to bind adaptors: two-hybrid
technology, in vitro binding assays, localization analysis, and
IP of TAC chimeras.
Example
The Lamp2 YEQF motif fulfills the criteria to be a signal recog-
nized by AP3.
3.6 Back-to-Context Mutation (truncations followed by point mutations) of candidate

Analysis signal within native context followed by adaptor-cargo IP and
localization analysis.
4 Conclusion
This chapter discusses the basis and application of a simple method

for the identification/testing of candidate sorting signals in TM
proteins. The understanding and control of protein sorting and
organization is not only important to further our basic knowledge
of cell function but also to support the development of medicinal
and biotechnological applications.
The proposed protocol takes advantage of the knowledge
obtained by researchers using the genetics, molecular biology, and
biochemistry toolkits. Now is the turn of information and computa-
tional approaches to move our knowledge and capabilities forward.
These approaches will not only create tools to facilitate the analysis
of protein information, but will also allow the generation of new
hypotheses to be tested on the classical experimental realm.
4.1 Where The method described in this chapter has been successfully and
to Go from Here? routinely used to analyze protein cargo of different origins (yeast,
mammalian, etc.). However, in its current form, it relies on manual
inspection of sequences and on the use of scattered resources. One
obvious limitation of this “manual” approach is its inability to sup-
port high-throughput analysis. Although complete collections of
protein sequences from entire organisms exist, we lack the tools to
systematically analyze the databases for cargos carrying specific
sorting signals. This limitation, for example, precludes us from
searching for cargoes potentially affected by genetic diseases due to
deficiencies in specific adaptors (e.g., Hermansky-Pudlak syndrome
forms due to AP3 abnormal function—see [15]). If we had this
capability, we could predict (and perhaps prevent) cellular abnor-
malities in patients.
Therefore, there is a clear need for applications that incorpo-
rate the myriad of factors discussed in previous sections.
Furthermore, the application of information theory [16] should be
instrumental for the development of methods to crack the protein-
sorting information code.
4.2 Is Benchwork Absolutely. Our basic cell and molecular knowledge still needs to
Still Needed? be pushed forward and classical benchwork will play a central role
in this effort. Indeed, there still are some basic research aspects that
need to be resolved/clarified, for example, controversies about the
specificity of certain adaptors and several reports of functional sig-
nals that deviate from the discussed consensuses. The existence of
previously unrecognized signals in the cytoplasm region of cargoes
is another possibility that should be experimentally investigated.
In addition to discussing the basis and application of a method
based on resources and analytical tools, from molecular biology to
ANNs, this chapter emphasizes the necessity to amalgamate exper-
imental and theoretical methods. The coalescence of in silico,
in vitro, and in vivo approaches is what it will take to fully under-
stand adaptor-signal interaction and to therefore move forward
knowledge and to fulfill our medical and biotechnological needs.
5 Notes
1. Emergent properties: Properties of the whole that do not

result from the sum of the parts. Examples: based on the
known properties of proteins, lipids, sugars, and nucleic acids,
it is unlikely that we would predict that their combination
would lead to a living cell. At the multicellular level, the
properties of individual neurons and accessory cells do not

predict that their grouping would lead to consciousness.
2. For basic concepts and terminology in cell biology and bio-
chemistry, from the 1- and 3-letter amino acid codes to the
mechanisms of vesicle trafficking, a few textbooks are
recommended:
Bruce Alberts, Alexander Johnson, Julian Lewis, Martin
Raff, Keith Roberts, Peter Walter (2007). Molecular Biology
of the Cell, 5th edition, Garland Science. ISBN: 0815341059;
ISBN-13: 9780815341055.
Thomas D. Pollard & William C. Earnshaw (2008). Cell
Biology, 2nd edition; Saunders/Elsevier. ISBN10:1416022554;
ISBN13: 9781416022558.
3. Integral membrane proteins: These proteins are partially
embedded into and permanently attached to the membranes
of cellular compartments. Transmembrane (TM) proteins
(TMP) are a very large subset of integral membrane proteins
and are characterized by spanning the membrane. TMP have a
TM domain or region made up of approximately two dozen
hydrophobic amino acids that separate lumenal and cytosolic
regions (see Fig. 2).
4. The effect on protein structure/function of amino acids such
as aspartic and glutamic acids (i.e., with a negatively charged
carboxyl group) can be emulated by amino acids serine and
threonine following the introduction of a negatively charged
phosphate group by phosphorylation. In fact, in order to test
the functional relevance of their phosphorylation, amino acids
serine and threonine are often intentionally mutated to aspar-
tic or glutamic acid.
5. However, several domains or 3D structures (e.g., ubiquitin)
have been found to be capable of conveying sorting informa-
tion on their 3D array of exposed amino acids (conformational
or 3D motifs).
6. Introduction to Biochemistry and cell biology techniques
(e.g., immunoprecipitation, immunofluorescence) can be
found in Short Protocols in Cell Biology. Juan S. Bonifacino
(Editor), Mary Dasso (Editor), Joe B. Harford (Editor),
Jennifer Lippincott-Schwartz (Editor), Kenneth M. Yamada
(Editor). ISBN: 978-0-471-48339-7. February 2004.
Acknowledgments
We are indebted to Arpita Sen (UC Berkeley) and members of

the Aguilar lab for critical reading of the manuscript. The Aguilar
lab is supported by the Center for the Science of Information, an
NSF Science and Technology Center, under grant agreement

CCF-0939370, by the National Institutes of Health and by the
National Science Foundation under Grants No. 5 R21 CA151961-
02 and 1021377, respectively.
References
1. Esposito G, Clara FA, Verstreken P (2012) Golgi. Ann Bot 92(2):167–180. doi:10.1093/
Synaptic vesicle trafficking and Parkinson’s dis- aob/mcg134
ease. Dev Neurobiol 72(1):134–144. doi: 10. Barman S et al (2001) Transport of viral pro-
10.1002/dneu.20916 teins to the apical membranes and interac-
2. Aridor M, Hannan LA (2002) Traffic jams II: tion of matrix protein with glycoproteins in
an update of diseases of intracellular trans- the assembly of influenza viruses. Virus Res
port. Traffic 3(11):781–790. doi:10.1034/ 77(1):61–69. doi:10.1016/s0168-1702(01)
j.1600-0854.2002.31103.x 00266-0
3. Boehm M, Bonifacino JS (2002) Genetic anal- 11. Ohno H, Aguilar RC, Yeh D et al (1998) The
yses of adaptin function from yeast to mam- medium subunits of adaptor complexes recog-
mals. Gene 286(2):175–186. doi:10.1016/ nize distinct but overlapping sets of tyrosine-
s0378-1119(02)00422-5 based sorting signals. J Biol Chem 273(40):
4. Bonifacino JS (2014) Adaptor proteins involved 25915–25921. doi:10.1074/jbc.273.40.25915
in polarized sorting. J Cell Biol 204(1):7–17. 12. Aguilar RC, Boehm M, Gorshkova I et al (2001)
doi:10.1083/jcb.201310021 Signal-binding specificity of the mu 4 subunit of
5. Hirst J, Irving C, Borner GHH (2013) Adaptor the adaptor protein complex AP-4. J Biol
protein complexes AP-4 and AP-5: new players Chem 276(16):13145–13152. doi:10.1074/
in endosomal trafficking and progressive spastic jbc.M010591200
paraplegia. Traffic 14(2):153–164. doi:10.1111/ 13. Mukherjee D, Hanna CB, Aguilar RC (2012)
tra.12028 Artificial neural network for the prediction of
6. Ohno H (2006) Physiological roles of clathrin tyrosine-based sorting signal recognition by
adaptor AP complexes: lessons from mutant ani- adaptor complexes. J Biomed Biotechnol. doi:
mals. J Biochem 139(6):943–948. doi:10.1093/ 10.1155/2012/498031
jb.mvj120 14. Parrish JR, Gulyas KD, Finley RL Jr (2006)
7. Green N, Fang H, Kalies KU et al. (2001) Yeast two-hybrid contributions to interactome
Determining the topology of an integral mem- mapping. Curr Opin Biotechnol 17(4):387–
brane protein. Current protocols in cell biol- 393. doi:10.1016/j.copbio.2006.06.006
ogy/editorial board, Juan S Bonifacino [et al.]. 15. Dell’Angelica EC, Shotelersuk V, Aguilar RC
Chapter 5: Unit 5.2. doi: 10.1002/0471143030. et al (1999) Altered trafficking of lysosomal
cb0502s00 proteins in Hermansky-Pudlak syndrome due
8. Vergarajauregui S, Puertollano R (2006) Two to mutations in the beta 3A subunit of the
di-leucine motifs regulate trafficking of muco- AP-3 adaptor. Mol Cell 3(1):11–21.
lipin-1 to lysosomes. Traffic 7(3):337–353. doi:10.1016/s1097-2765(00)80170-7
doi:10.1111/j.1600-0854.2006.00387.x 16. Adami C (2012) The use of information theory
9. Neumann U, Brandizzi F, Hawes C (2003) in evolutionary biology. Year Evol Biol 1256:
Protein transport in plant cells: in and out of the 49–65.doi:10.1111/j.1749-6632.2011.06422.x
Chapter 2
Protein Structural Information Derived from NMR

Chemical Shift with the Neural Network Program TALOS-N
Yang Shen and Ad Bax
Abstract
Chemical shifts are obtained at the first stage of any protein structural study by NMR spectroscopy.
Chemical shifts are known to be impacted by a wide range of structural factors, and the artificial neural
network based TALOS-N program has been trained to extract backbone and side-chain torsion angles
from 1H, 15N, and 13C shifts. The program is quite robust and typically yields backbone torsion angles for
more than 90 % of the residues and side-chain χ1 rotamer information for about half of these, in addition
to reliably predicting secondary structure. The use of TALOS-N is illustrated for the protein DinI, and
torsion angles obtained by TALOS-N analysis from the measured chemical shifts of its backbone and 13Cβ
nuclei are compared to those seen in a prior, experimentally determined structure. The program is also
particularly useful for generating torsion angle restraints, which then can be used during standard NMR
protein structure calculations.
Key words NMR, Chemical shifts, Protein structure, Side-chain conformation, Artificial neural
network, Secondary structure, Backbone torsion angle
1 Introduction
1.1 Relations The first step of any protein structural study by NMR spectroscopy
Between Chemical typically involves assignment of the multitude of NMR resonances
Shifts and Protein to individual nuclei. Originally, for proteins extracted from natural
Structure sources, this only involved assignment of the hydrogen NMR
spectra [1, 2]. However, due to extensive resonance overlap in 1H
NMR spectra, this technology was restricted to relatively small
proteins. With advances in molecular biology, the vast majority of
today’s structural studies focus on cloned proteins, typically over-
expressed in Escherichia coli [3–5]. By using suitable isotopically
enriched growth media, it then is readily feasible to obtain essen-
tially full incorporation of the NMR-observable stable isotopes 13C
and 15N. These nuclei not only are key to dispersing the crowded
NMR spectra in three or four orthogonal frequency dimensions,
dramatically reducing the resonance overlap problem; the 13C and
17
18 Yang Shen and Ad Bax
15
N chemical shifts themselves have proven to be important reporters
on local backbone conformation [6–8]. NMR chemical shifts in
proteins are exquisitely sensitive to local conformation. However,
they depend on many different factors, including backbone and
side-chain torsion angles, neighboring residues, ring currents
caused by nearby aromatic groups, hydrogen bonding, electric
fields, local strain and geometric distortions, as well as solvent
exposure [9–15]. This not only has made it difficult to separately
quantify the relation between each of these parameters and the
chemical shift; it also makes it impossible to uniquely attribute
such a structural parameter to any individual chemical shift.
For protein NMR spectroscopy, triple resonance correlation
experiments, which link the resonances of directly bonded 1H, 13C,
and 15N nuclei, are commonly used to assign the chemical shifts of
1
H, 13C, and 15N nuclei in proteins [16–18]. The chemical shift
assignment procedure usually consists of two steps: (1) sequence-
specific assignment of the backbone atoms and (2) side-chain
assignments. Nearly complete chemical shift assignments for back-
bone and side-chain atoms are commonly required to assign
nuclear Overhauser enhancement (NOE) spectra, which classically
are used to derive interproton distances that serve as the primary
experimental restraints for calculating the protein structure. The
backbone (1Hα, 13C′, 13Cα, 15N, and 1HN) and 13Cβ chemical shifts,
which are generally obtained in the earliest stage of any protein
NMR study, are particularly useful reporters on local conforma-
tion. Their link to secondary structure, as well as to hydrogen
bonding and χ1 side-chain torsion angles, has been long recog-
nized and has been the focus of both empirical studies as well as
quantum-chemical calculations [11–15, 19, 20].
1.2 Protein The rapid increase in the number of proteins, for which both high-
Backbone and Side- resolution structural coordinates have been deposited in the
Chain Conformation Protein Data Bank (PDB) [21] and NMR chemical shift assign-
from NMR Chemical ments are available in the BioMagResBank (BMRB) [22], has
Shifts stimulated the development of quantitative empirical methods to
study the relation between protein structure and chemical shifts
[23]. Among the wide array of empirical methods, TALOS [20]
and its two successors TALOS+ [24] and TALOS-N [25] have
become particularly widely used for making accurate ϕ/ψ back-
bone torsion angle predictions on the basis of the backbone (13Cα,
13
C′, 15N, 1Hα, and 1HN) and 13Cβ chemical shift assignments.
These ϕ/ψ predictions can be used to validate NOE-derived NMR
structures that did not use chemical shift-derived input parameters
or, conversely, to generate additional restraints as input to the pro-
tein structure calculation and refinement protocols.
The original TALOS program (Torsion Angle Likeliness
Obtained from Shift) searches a protein database, consisting origi-
nally of only 20 proteins but later expanded to ca 200 proteins,
Protein Structure from NMR 19
with both high-resolution X-ray coordinates and NMR chemical

shift assignments. TALOS identifies the ten tripeptide fragments
that represent the best match in terms of chemical shifts and resi-
due types to those of a tripeptide segment whose assignments are
known and whose structure is under study (the “target protein”).
The assumption underlying TALOS is that fragments with similar
chemical shifts and residue type typically have similar backbone
conformations. Thus, if these ten best-matched fragments have
consistent, narrowly clustered values for the ϕ/ψ angles of their
center residue, their averages and standard deviations are used as a
prediction for the ϕ/ψ angles of the center residue of the target
protein tripeptide. If the ϕ/ψ angles of the center residue of these
ten best-matched tripeptides fall in different regions of the
Ramachandran map, the matches are declared ambiguous, and no
prediction is made for the central residue. With this quality control
criterion, TALOS predicted ϕ/ψ torsion angles for, on average, ca
72 % of the residues in any given target protein. For TALOS valida-
tion proteins, where the true ϕ/ψ angles are known, only about
1.8 % of the predictions were inconsistent with crystallographically
determined ϕ/ψ torsion angles. Excluding these 1.8 % erroneous
predictions, a root mean square (RMS) difference of ca 13° is
observed between predicted and crystallographically observed ϕ/ψ
torsion angles.
Although rather robust, the original TALOS program was
unable to make definitive predictions for about 28 % of the resi-
dues in any given protein. Most of these 28 % are located outside
regular secondary structure, exactly those regions where backbone
torsion angle information is most needed. TALOS+ was developed
to address this shortcoming and to extend the coverage of the pro-
gram [24]. For a given residue in the target protein, TALOS+ first
uses an artificial neural network (ANN) module to predict its
three-state distribution in the Ramachandran map, i.e., α, β, and
positive-ϕ. This three-state distribution is subsequently used to
guide the database search procedure for the ten best matches. With
the incorporation of the ANN, TALOS+ is able to increase its cov-
erage to ca 88 %, without sacrificing accuracy. Thus, compared to
the original TALOS program, the fraction of residues whose back-
bone angles cannot be predicted is reduced from ~28 % to ~12 %.
Importantly, most of the additional ϕ/ψ torsion angle predictions
are made for residues in loop or turn regions, where this informa-
tion is needed most.
The recently introduced TALOS-N program relies far more
extensively on neural network analysis of the input chemical shift data
than TALOS+, thereby further increasing coverage, accuracy, and
reliability. In addition, TALOS-N is the first program to g enerate
quite accurate predictions for the side-chain χ1 torsion angles (Fig. 1).
For the ϕ/ψ torsion angle prediction of a given residue i in the
target protein, a well-trained two-level feed-forward multilayer
Fig. 1 Flowchart of the TALOS-N program (reproduced from [25] with permission from Springer)
ANN, referred to as a (ϕ,ψ)-ANN, is first used by TALOS-N to

predict the 324-state ϕ/ψ distribution of residue i on the basis of
the NMR chemical shifts and residue type of itself and its adjacent
residues (i − 2 to i + 2). Here, the 324-state ϕ/ψ distribution cor-
responds to the likelihood that residue i adopts torsion angles that
fall in any of the 324 voxels, of 20° × 20° each, that make up the
Ramachandran map. The ANN-predicted ϕ/ψ distribution is then
used solely to search a large crystallographic database (containing
9,523 proteins, with chemical shifts added by a computational
method [26]), for a pool of 1,000 heptapeptide fragments with
ϕ/ψ angles that best match the 324-state ϕ/ψ distribution. These
top 1,000 fragments then are further evaluated for the agreement
between their computed chemical shifts and experimental values of
the corresponding heptapeptide segment (i − 3 to i + 3) in the tar-
get protein. The 25 best-matched database heptapeptides are
retained, and the ϕ/ψ angles of their center residues are inspected
by using an advanced clustering analysis, and subsequently used
to make a prediction for the ϕ/ψ angles of the query residue.
Validation on an independent set of proteins indicates that backbone
torsion angles can be predicted for a larger, ≥90 % fraction of the

residues, with an error rate of ca 3.5 % when using an acceptance
criterion that is nearly twofold tighter than that used previously by
TALOS and TALOS+. The RMS difference between predicted and
crystallographically observed ϕ/ψ torsion angles is ca 12°, also
slightly better than what was obtained with the earlier versions of
the program.
To predict the χ1 rotameric state (g−, g+ or t) for a given residue
i (of residue type a) in the target protein, TALOS-N uses another
set of ANNs, referred to as (χ1)a-ANNs. The (χ1)a-ANN has been
trained to correlate the center residue likelihood of adopting each
of the three χ1 rotameric states to the differences between its
observed chemical shifts and those expected on the basis of its
backbone conformation. A separate database search procedure is
subsequently used to estimate the three-state probability of residue
i to adopt the three χ1 rotameric states. With an optimized error
control criterion, TALOS-N predicts χ1 rotameric states for ca
50 % of the residues, with an “error rate” of ca 10 % when compar-
ing the predicted χ1 rotameric state to that of any given reference
structure. However, we note that the true error is likely to be much
lower, as for proteins that have multiple available independently
solved X-ray structures, the χ1 rotameric states of any “erroneous”
χ1 prediction is typically in agreement with that of another X-ray
structure [25].
Similar to TALOS+, TALOS-N is also implemented with an
ANN-based module for predicting secondary structure (SS) from
the NMR chemical shifts. For this purpose, TALOS-N uses two
separate ANNs, referred to as SS-ANN and SSseq-ANN, which are
trained to correlate the three-state secondary structure classifica-
tion (helix, sheet, and coil) of a residue to both the chemical shifts
and amino acid sequence or to amino acid sequence alone, respec-
tively. The output of these two ANNs is used in a hybrid manner
to predict secondary structure for any residue in a protein, regard-
less of the completeness of chemical shift assignments. The overall
correctness of the SS prediction is ca 88 % when NMR chemical
shifts are available, dropping to ca 81 % when no chemical shifts are
available. In the absence of chemical shifts, TALOS-N matches the
accuracy of the best sequence-only secondary structure prediction
programs [27, 28].
2 Materials
In this chapter, we use the protein DinI [29] to illustrate the use of
TALOS-N for predicting its backbone ϕ/ψ and side-chain χ1 torsion
angles, as well as its secondary structure classification. To follow the
examples, both the TALOS-N software package and an input file
with correctly formatted chemical shift assignments are needed.
2.1 Software The TALOS-N software package, including the required binaries
Requirements for three of the most common operating systems, Linux, Mac OS
X, and Windows, as well as the requisite protein database and
scripts, can be downloaded from http://spin.niddk.nih.gov/bax/
software/TALOS-N/ and installed straightforwardly (see Note 1).
Alternatively, a server version of TALOS-N can be used directly,
without installing the TALOS-N software (http://spin.niddk.nih.
gov/bax/nmrserver/talosn/).
2.2 Data An input table containing both the full amino acid sequence and
Requirements the NMR chemical shift assignments is required, to be prepared
with a specific data format (general purpose NMRPipe table for-
mat). As an example, an excerpt of such a file is shown below for
the protein DinI:
DATA FIRST_RESID 2
DATA SEQUENCE RIEVTIAKT SPLPAGAIDA LAGELSRRIQ
YAFPDNEGHV SVRYAAANNL
DATA SEQUENCE SVIGATKEDK QRISEILQET WESADDWFVS E
VARS RESID RESNAME ATOMNAME SHIFT
FORMAT %4d %1s %4s %8.3f
2 R C 174.123
2 R CA 55.537
2 R CB 32.786
2 R H 8.772
2 R HA 4.994
2 R N 123.394
3 I C 173.941
3 I CA 60.986
The protein’s amino acid sequence should be provided in one
or more lines starting with the tag “DATA SEQUENCE”. Only the
one-character residue name is allowed (see Note 2) and space char-
acters in the sequence are ignored. An optional line beginning with
a tag of “DATA FIRST_RESID” is needed to specify the first resi-
due number of the amino acid sequence listed in the “DATA
SEQUENCE” line if the first residue listed is not residue number 1.
For the chemical shift table, columns for residue number, one-
character residue type (see Note 2), atom name (see Note 3), and
chemical shift value must be included, and their definitions
(“RESID,” “RESNAME,” “ATOMNAME,” and “SHIFT,” respec-
tively) must be predeclared in a line beginning with a “VARS” tag;
a line beginning with a “FORMAT” tag is also required (immedi-
ately after the “VARS” line) to define the data type of each corre-
sponding column of the table.
Note that all chemical shifts used as input for TALOS-N are
required to be properly referenced (see Note 4) to ensure the accu-
racy and reliability of the predictions. If the protein sample used to
collect the NMR chemical shift data is perdeuterated, 2H isotope
corrections [30] need to be applied for 13Cα and 13Cβ chemical

shifts (see Note 5).
Other standard chemical shift formats, such as the NMR-Star
format used by the BMRB database, can also be used as input after
a format conversion. A conversion script is provided in the
TALOS-N package for this purpose (see Note 6). The server ver-
sion of TALOS-N includes automated chemical shift format iden-
tification and can use the NMR-Star format chemical shift file
directly as input, without requiring prior format conversion.
3 Methods
3.1 TALOS-N The TALOS-N prediction can be performed for DinI with an input
Prediction chemical shift file of name “inCS.tab” by typing the command:
talosn -in inCS.tab
The program first converts the chemical shifts (δ) of each query
residue to its corresponding secondary chemical shifts (Δδ) by sub-
tracting a residue-type-dependent random coil value, as well as
corrections to account for the residue types of its two immediate
neighbors. The converted secondary chemical shifts are stored in a
file named “predAdjCS.tab” (in the “SHIFT” column),
together with the original chemical shifts (“CS_OBS”) and the cor-
responding corrections (“CS_ADJ”, which is the random coil value
including nearest neighbor (i ± 1) residue-type correction) used to
calculate the secondary chemical shifts. To make a ϕ/ψ angle pre-
diction, the converted secondary chemical shifts together with the
amino acid-type information are used as inputs for the (ϕ/ψ)-ANN
to calculate the 324-state ϕ/ψ distribution for each predictable
residue (see Note 7), with the output stored in a file named “pre-
dANN.tab”. A database search step is then performed to search a
9,523-protein database for the 25 best-matched heptapeptides in
terms of the 324-state ϕ/ψ angle distribution, the secondary
chemical shifts, and the amino acid type. A single file, “predAll.
tab”, is generated in this step to store the information of those
best database matches for each of the residues in the target protein.
A final summarization and quality control step is performed to
identify outliers in the 25 best-matching heptapeptides by evaluat-
ing the clustering of the ϕ/ψ angles of their center residues in the
Ramachandran map or by using the observed ϕ/ψ of a reference
structure if such a structure is available (this requires an additional
option “-ref ref.pdb” in the command line, where “ref.
pdb” is the name for the reference structure). A summary file
“pred.tab” is then created, displaying the average ϕ and ψ values
(in the PHI and PSI columns) and their respective standard devia-
tions (DPHI and DPSI), as well as an aggregate, weighted χ2 score
(DIST, see Eq. 12 of reference [25]), reflecting how well the target
protein chemical shifts match those of the database fragments.

An excerpt of this file for DinI is shown below:
VARS RESID RESNAME PHI PSI DPHI DPSI DIST
S2 COUNT CS_COUNT CLASS
FORMAT %4d %s %8.3f %8.3f %8.3f %8.3f %8.3f
%5.3f %2d %2d %s
2 R -107.206 129.115 9.502 7.843
0.293 0.873 25 16 Strong
3 I -117.237 126.352 6.691 6.523
0.180 0.883 25 18 Strong
For each predictable residue, or residue with sufficient input
chemical shifts (see Note 7), a final classification is made (listed as
the last “CLASS” column in the “pred.tab” file) for its ϕ/ψ
angle prediction by a summarization step, detailed below. Prior to
making this final classification, the program calculates the predicted
backbone rigidity as reflected in the “random coil index” order
parameter, RCI-S2 [31], which scales between 0 (total disorder)
and 1 (fully rigid). Its values are included under the “S2” column
in the “pred.tab” file. Residues below the threshold RCI-S2
≤0.6 are assigned as dynamic (receiving a “Dyn” classification) in
“pred.tab”. For other residues, a classification of strongly unam-
biguous is assigned (with a “Strong” tag) if the center residues of
all 25 best-matching heptapeptides locate in a consistent ϕ/ψ
region in the Ramachandran map. A generously unambiguous
classification is assigned (with a “Generous” tag) if the center
residues of only the top ten best matches cluster in a consistent
ϕ/ψ region. All other cases are considered ambiguous (classified
with a “Warn” tag), even though inspection of their Ramachandran
map population may contain very useful information. For example,
often the ambiguous residues will cluster in two distinct regions of
the Ramachandran map, and the investigator can explore both
options during structure calculations.
For the predictable residues, the ϕ/ψ angles are calculated by
averaging the ϕ/ψ angles of the center residues of all 25 best
matches (for residues classified as “Strong”) or from the top ten
best matches (for a “Generous” prediction) and shown in the
“PHI”/“PSI” columns. The estimated uncertainties in the
predicted ϕ/ψ angles are calculated from their standard deviations
from these averages and listed in the “DPHI” and “DPSI” col-
umns. Only when a known reference structure is provided as input
to the program will the predicted ϕ/ψ values be compared to the
observed ϕ/ψ angles in this reference structure for all unambigu-
ously predicted (“Strong” and “Generous”) residues. A predic-
tion is labeled as “Bad” if the predicted and the observed ϕ/ψ
angles are not consistent (see Note 8).
For DinI, 71 residues (out of a total of 81) are obtained
with unambiguous ϕ/ψ angles prediction, 2 have an ambiguous
ϕ/ψ angle prediction, and 6 are predicted as dynamic (Fig. 2).
Fig. 2 Graphic TALOS-N inspection interface for protein DinI. (For details, see Subheading 3.2 or the TALOS-N
webpage http://spin.niddk.nih.gov/bax/software/TALOS-N/)
Among those 71 unambiguous predictions, 70 are classified as

“Strong” and two (Ala15 and Gly16) are designated “Bad” after
inspecting their consistency relative to the ϕ/ψ angles observed in
the reference NMR structure. It is worth noting that, in the refer-
ence structure, these latter two residues (with ϕ/ψ angles of
−57°/49° and −176°/−18°, respectively) are located in very lowly
populated regions of the Ramachandran map, i.e., they are statisti-
cally unlikely to occur. Without further experimental data, it is not
possible to decide whether the “Bad” classification refers to the
reference structure or to the quality of the prediction.
After TALOS-N prediction of ϕ/ψ angles has been completed,
another database search and ANN-based procedure is performed
to predict the χ1 rotameric states. A χ1 rotamer prediction summary
file “predChi1.tab” is created with an excerpt of this file shown
below for DinI:
VARS RESID RESNAME CS_COUNT CHI1_OBS Q_Gm Q_Gp Q_T CLASS
FORMAT %4d %s %2d %8.3f %5.3f %5.3f %5.3f %s
2 R 16 -69.938 0.341 0.121 0.538 na
3 I 18 -62.494 0.873 0.063 0.063 g-
4 E 18 -61.087 0.312 0.093 0.595 na
5 V 18 178.554 0.073 0.055 0.872 t
6 T 18 66.182 0.302 0.464 0.235 na
7 I 18 -75.725 0.713 0.143 0.143 g-
For a query residue of residue type a (excluding Gly, Pro, and

Ala) with sufficient input chemical shifts (see Note 7), TALOS-N
first searches the database for the 1,000 best-matched heptapep-
tides in terms of the backbone torsion angles and residue types.
It then uses a trained (χ1)a-ANN to calculate a χ1 matching score
for each database match, which measures the likelihood of the
center residue of the database heptapeptide to adopt the same χ1
rotameric state as the query residue. The program then derives a
three-state probability score, Pc, for the query residue to adopt
each of the three χ1 rotameric states (c = g−, g+, and t, stored in the
columns “Q_Gm,” “Q_Gp,” and “Q_T,” respectively, in “pred-
Chi1.tab”). TALOS-N then classifies the prediction for the
query residue to adopt χ1-rotamer state c (g−, g+, or t, as listed in
the last column of “CLASS” in the “predChi1.tab” file) only
when the predicted probability for state c is significantly higher
than that for the other two states, by default Pc > 0.6. Otherwise, an
ambiguous classification is assigned (with a “na” tag). Details of
other contents in “predChi1.tab” are as follows: the column of
“CS_COUNT” is for the count of the experimental chemical shifts
of the target residue itself and its two neighbors; when a reference
structure is provided, a “CHI1_OBS” column is provided to dis-
play the χ1 angle observed in the reference structure. For DinI,
TALOS-N makes χ1 rotameric state predictions for 30 out of a
total of 61 (non-Gly/-Pro/-Ala) residues, among which three
(Asp35, Asn48, and Asp75) differ in their predicted χ1 rotameric
state from the reference NMR structure (PDB entry 1GHH).
Next to predicting ϕ, ψ, and χ1 torsion angles, TALOS-N also
predicts the protein’s secondary structure. For residues with chem-
ical shift assignments, a two-level neural network, SS-ANN, is
trained to make a three-state secondary structure prediction (H, E,
or L, representing for α-helix, β-sheet, and loop, respectively) on
the basis of both the chemical shifts and the amino acid sequence
information. In addition, another two-level ANN, referred to as
SSseq-ANN, is trained by using solely the amino acid sequence
information. It can be used to make predictions for residues that
lack chemical shift information. However, this SSseq-ANN is used
more generally by TALOS-N in a hybrid manner with the SS-ANN
to make secondary structure prediction for proteins when chemical
shift assignments are incomplete. TALOS-N generates an output
file “predSS.tab” to store the predicted secondary structure. An
excerpt of this file is shown below for DinI:
VARS RESID RESNAME CS_CNT CS_CNT_R2 Q_H Q_E Q_L CONFIDENCE SS_CLASS
FORMAT %4d %1s %2d %2d %8.3f %8.3f %8.3f %4.2f %s
1 M 10 4 0.333 0.333 0.333 0.00 L
2 R 16 6 0.097 0.740 0.162 0.58 E
3 I 18 6 0.027 0.970 0.003 0.94 E
4 E 18 6 0.006 0.968 0.026 0.94 E
5 V 18 6 0.004 0.963 0.033 0.93 E

6 T 18 6 0.009 0.970 0.021 0.95 E
Details of its contents are as follows: the column of “CS_CNT_
R2” lists the number of experimental chemical shifts of the target
residue; “CS_CNT” contains the count of experimental chemical
shifts of the target residue plus its two immediate neighbors; the
columns “Q_H,” “Q_E,” and “Q_L” list the SS-ANN (or SSseq-
ANN) predicted probability for the target residue to be of second-
ary structure type “H,” “E,” and “L,” respectively; the values
shown in the “CONFIDENCE” column represent the confidence of
the three-state secondary structure prediction for a given target
residue, calculated from the difference of maximal and median val-
ues of “Q_H,” “Q_E,” and “Q_L”; and the text listed in the “SS_
CLASS” column shows the final secondary structure classification
assigned by the program, i.e., one of the three states with the maxi-
mal predicted probability.
For DinI, when comparing to the output of the DSSP pro-
gram [32] for the reference structure (PDB entry 1GHH), the
overall correctness ratio of the TALOS-N predicted secondary
structure is 70/81. In this respect, it is important to note that,
even for proteins of known structure, secondary structure assign-
ment can be ambiguous, as reflected in only ca 90 % agreement
among the output of some of the most popular programs [23].
As mentioned above, TALOS-N predictions can either be
made locally by downloading the requisite programs or be per-
formed via the TALOS-N server (http://spin.niddk.nih.gov/bax/
nmrserver/talosn/), which requires a chemical shift file as input
and an e-mail address to send back the prediction results, including
all abovementioned output files, such as “pred.tab”, “predChi1.
tab”, “predSS.tab”, “predS2.tab”, “predAll.tab”,
“predAdjCS.tab”, and “predANN.tab”.
3.2 Manual The TALOS-N predictions can be inspected and further adjusted
Inspection by using a Java graphic program, jrama. Two examples of com-
and Adjustment mand line calls of this program are:
jrama -in pred.tab
jrama -in pred.tab -ref DinI.pdb
Figure 2 shows the jrama graphic interface, loaded with the
TALOS-N predicted results for DinI. The left panel of the graphic
interface shows a map of the ϕ/ψ angles of the center residues of
the 25 best-matched heptapeptides in the database (green
squares) and the query residue Thr-6 (blue, depicting the angles
observed in the NMR-derived PDB entry 1GHH), superimposed
on a Ramachandran map, depicting in gray the “most favorable”
ϕ/ψ angles for Thr, i.e., those most commonly observed in high-
resolution crystal structures of a very large array of proteins. The
324 (ϕ/ψ)-ANN-predicted scores for Thr-6 are shown as colored
voxels but only for those that are populated at least one standard
deviation above the average predicted voxel density. The top
right panel displays the amino acid sequence of DinI, with resi-
dues colored according to their ϕ/ψ prediction classification.
Missing predictions (e.g., residue M1) are shown in light gray,
consistent predictions in light or dark green (for “Strong” and
“Generous” predictions, respectively), ambiguous predictions
in yellow, and dynamic residues in blue. Three other panels cor-
respond to the RCI-S2 value, the predicted secondary structure
(red, α-helix; aqua, β-sheet), with the height of the bars reflecting
the probability assigned by the SS-ANN secondary structure pre-
diction. The bottom right panel depicts the χ1 rotamer predic-
tions (red oval, g−; green, g+; yellow, t), with the height of the
ovals corresponding to the probability assigned by the χ1 rota-
meric state prediction.
The TALOS-N prediction (including the summary of
TALOS-N predicted ϕ/ψ angles) is normally performed with the
default parameters and settings. However, the left panel also can
be used to manually adjust the prediction classification of a given
query residue according to a user’s preference. The prediction
files then will be overwritten to reflect any changes made
interactively.
3.3 Generation The TALOS-N output can be converted into ϕ and ψ torsion
of Angular Restraints angle restraints that then can be used directly as input for a con-
ventional protein NMR structure calculation procedure [33, 34].
Two convenient scripts, “talos2dyana.com” and “talos2x-
plor.com”, are included in the TALOS-N software package for
this purpose. These scripts read predicted ϕ and ψ angles from
the TALOS-N prediction summary file “pred.tab” and gener-
ate for each residue with an unambiguous TALOS-N prediction
(classified as “Strong” or “Generous”) a ϕ and a ψ torsion
angle restraint (see Note 9). These torsion angle restraints can be
stored in either CYANA format, as shown below for residues 2
and 3 of DinI:
2 ARG PHI -127.2 -87.2
2 ARG PSI 109.1 149.1
3 ILE PHI -137.2 -97.2
3 ILE PSI 106.4 146.4
or in XPLOR format:
assign (resid 1 and name C )(resid 2 and name N )
(resid 2 and name CA )(resid 2 and name C ) 1.0 -107.2 20.0 2
assign (resid 2 and name N ) (resid 2 and name CA )
(resid 2 and name C ) (resid 3 and name N ) 1.0 129.1 20.0 2
assign (resid 2 and name C ) (resid 3 and name N )
(resid 3 and name CA ) (resid 3 and name C ) 1.0 -117.2 20.0 2
assign (resid 3 and name N ) (resid 3 and name CA )
(resid 3 and name C ) (resid 4 and name N ) 1.0 126.4 20.0 2
These input restraints then can be used for protein structure

calculations as a complement to the conventional NOE distance
restraints. Note that such chemical shift-derived torsion angle
restraints alone are typically insufficient to reach a converged pro-
tein structure as each torsion angle contains a substantial uncer-
tainty (±20°, in the above example), and these uncertainties rapidly
accumulate when building the protein chain. Moreover, as men-
tioned above, predictions are generally only about 90 % complete
and may contain errors.
4 Notes
1. An installation shell script “install.com” is provided with

the TALOS-N software package, which can be use for installing
and configuring the TALOS-N program on a Linux or a Mac
OS X system. After the installation, two starting shell scripts
“talosn” and “jrama” are generated with properly config-
ured installation paths for the system-specific binary and all
required databases. For a Windows system, TALOS-N can be
installed by simply uncompressing the package. However, when
running the TALOS-N program, the TALOS-N installation
path (“$talosnDir”) must be specified on the fly, for example,
with the command (see Subheading 3.1) of “$talosnDir/
bin/TALOS.exe –in inCS –talosnDir $talosnDir”.
2. In both the sequence header and the chemical shift data table,
the lowercase “c” must be used for oxidized Cys (δ13Cβ ~
42.5 ppm) and uppercase “C” for reduced Cys (δ13Cβ ~ 28 ppm),
“h” for protonated His, and “H” for deprotonated His.
3. Atom names should be given exactly as: “HA” for Hα atoms of
all non-Gly residues; “HA2” for the first Hα atom of a Gly resi-
dues and “HA3” for the second; “C” for C′ (CO) atoms; “CA”
for Cα atoms; “CB” for Cβ atoms; “N” for amide nitrogen
atoms; and “HN” for amide hydrogens. Data for all other atom
types will be ignored.
4. All 13C chemical shifts (including δ13Cα, δ13Cβ, and δ13C′)
should be referenced relative to the methyl groups of
4,4-dimethyl-4-silapentane-1-sulfonic acid, or DSS [35]. The
15
N chemical shifts used as input for TALOS-N should be ref-
erenced relative to liquid ammonia at 25 °C [35]. A pre-check
module in TALOS-N will be used to identify possible referencing
problems with the δ13Cα, δ13Cβ, δ13C′, and δ1Hα chemical shift
inputs [36] when running a typical TALOS-N command with
an additional “-check” option, for example, by using the
command line input argument “talosn -in inCS.tab
-check”. This module first converts the chemical shifts (δ) of
each residue to secondary chemical shifts (Δδ; see
Subheading 3.1) and subsequently evaluates these by correlat-

ing Δδ13Cα, Δδ13Cβ, Δδ13C′, and Δδ1Hα to the reference-free
entity, Δδ13Cα − Δδ13Cβ [36]. The estimated chemical shift ref-
erencing offsets, as well as their corresponding fitting error,
will be printed for δ13Cα, δ13Cβ, δ13C′, and δ1Hα. An offset
correction generally is only needed when the estimated refer-
encing offset exceeds the average fitting error by more than
about five standard deviations. This pre-check module will also
identify residues with unusual chemical shifts, for which sec-
ondary chemical shifts fall outside the expected range. Such
chemical shift outliers, especially those with highly unusual
chemical shifts, for which secondary chemical shifts deviate
from the expected range by more than two times of the normal
range of secondary chemical shifts, may correspond to experi-
mental errors and need to be inspected carefully prior to using
them for making torsion angle predictions.
5. 2H isotope chemical shift corrections for 13Cα and 13Cβ [30]
can be applied before starting the TALOS-N prediction, i.e.,
when generating the secondary chemical shifts. To do this, an
additional option “-iso” must be added when running a
TALOS-N prediction, for example, by using a command line
argument of the form “talosn -in inCS.tab -iso”.
6. A conversion Unix shell script, bmrb2talos.com, is included
with the TALOS-N package and can be used to convert a
NMR-Star format chemical shift table, used by the BMRB data-
base, to TALOS format. An example command line for using
this script is “bmrb2talos.com bmrb.str > inCS.tab”.
7. To ensure the prediction accuracy and reliability for a given
query residue, the chemical shift sufficiency is first inspected by
the program for the residue itself and its two immediate neigh-
bors. If at least two of the three residues have at least three
chemical shifts, the center residue is considered to be
predictable.
8. The consistency between the predicted ϕ/ψ values (ϕpred/ψpred)
and the observed ϕ/ψ angles (ϕobs/ψobs) is defined by
(f - fobs ) + (y pred - y obs ) < 60° .

2 2
pred
9. For a residue with a “Strong” classification of its prediction,

the ϕ and ψ angle restraints are set to <ϕ> ± 2σ and <ψ> ± 2σ,
where <φ> and <ψ> are the averaged TALOS-N predictions
and 2σ is the larger of 20° or two standard deviations of the
TALOS-N prediction. For a residue classified with a
“Generous” prediction, the ϕ and ψ angle restraints are less
tight, <ϕ> ± 3σ and <ψ> ± 3σ, with an allowed range of the
larger of 30° or three standard deviations of the TALOS-N
prediction.
Acknowledgments
This work was funded by the Intramural Research Program of the

NIDDK, NIH.
References
1. Wüthrich K (1986) NMR of proteins and 13. Vila JA, Aramini JM, Rossi P et al (2008)
nucleic acids. Wiley, New York Quantum chemical C-13(alpha) chemical shift
2. Englander SW, Wand AJ (1987) Main-chain- calculations for protein NMR structure deter-
directed strategy for the assignment of 1H mination, refinement, and validation. Proc Natl
NMR spectra of proteins. Biochemistry 26: Acad Sci U S A 105:14389–14394
5953–5958 14. Kohlhoff KJ, Robustelli P, Cavalli A et al
3. Oh BH, Westler WM, Darba P et al (1988) (2009) Fast and accurate predictions of protein
Protein 13C spin systems by a single two- NMR chemical shifts from interatomic dis-
dimensional nuclear magnetic resonance exper- tances. J Am Chem Soc 131:13894–13895
iment. Science 240:908–911 15. Asakura T, Taoka K, Demura M et al (1995)
4. Ikura M, Kay LE, Bax A (1990) A novel The relationship between amide proton chemi-
approach for sequential assignment of 1H, 13C, cal shifts and secondary structure in proteins. J
and 15N spectra of larger proteins: heteronu- Biomol NMR 6:227–236
clear triple-resonance three-dimensional NMR 16. Bax A, Grzesiek S (1993) Methodological
spectroscopy. Application to calmodulin. advances in protein NMR. Acc Chem Res
Biochemistry 29:4659–4667 26:131–138
5. Wagner G (1993) Prospects for NMR of large 17. Sattler M, Schleucher J, Griesinger C (1999)
proteins. J Biomol NMR 3:375–385 Heteronuclear multidimensional NMR exper-
6. Saito H (1986) Conformation-dependent C13 iments for the structure determination of
chemical shifts—a new means of conforma- proteins in solution employing pulsed field
tional characterization as obtained by high res- gradients. Prog Nucl Magn Reson Spectrosc
olution solid state C13 NMR. Magn Reson 34:93–158
Chem 24:835–852 18. Salzmann M, Wider G, Pervushin K et al
7. Wishart DS, Sykes BD, Richards FM (1991) (1999) TROSY-type triple-resonance experi-
Relationship between nuclear magnetic reso- ments for sequential NMR assignments of large
nance chemical shift and protein secondary proteins. J Am Chem Soc 121:844–848
structure. J Mol Biol 222:311–333 19. Wagner G, Pardi A, Wuthrich K (1983)
8. Spera S, Bax A (1991) Empirical correlation Hydrogen-bond length and H-1-NMR
between protein backbone conformation and chemical- shifts in proteins. J Am Chem Soc
Ca and Cb 13C nuclear magnetic resonance 105:5948–5949
chemical shifts. J Am Chem Soc 113: 20. Cornilescu G, Delaglio F, Bax A (1999) Protein
5490–5492 backbone angle restraints from searching a
9. Haigh CW, Mallion RB (1979) Ring current database for chemical shift and sequence
theories in nuclear magnetic resonance. Prog homology. J Biomol NMR 13:289–302
Nucl Magn Reson Spectrosc 13:303–344 21. Berman HM, Kleywegt GJ, Nakamura H et al
10. Avbelj F, Kocjan D, Baldwin RL (2004) (2012) The Protein Data Bank at 40: reflecting
Protein chemical shifts arising from alpha- on the past to prepare for the future. Structure
helices and beta-sheets depend on solvent 20:391–396
exposure. Proc Natl Acad Sci U S A 101: 22. Markley JL, Ulrich EL, Berman HM et al
17394–17397 (2008) BioMagResBank (BMRB) as a partner
11. de Dios AC, Pearson JG, Oldfield E (1993) in the Worldwide Protein Data Bank (wwPDB):
Secondary and tertiary structural effects on new policies affecting biomolecular NMR
protein NMR chemical shifts—an ab initio depositions. J Biomol NMR 40:153–155
approach. Science 260:1491–1496 23. Wishart DS (2011) Interpreting protein chem-
12. Case DA (1998) The use of chemical shifts and ical shift data. Prog Nucl Magn Reson Spectrosc
their anisotropies in biomolecular structure 58:62–87
determination. Curr Opin Struct Biol 8: 24. Shen Y, Delaglio F, Cornilescu G et al (2009)
624–630 TALOS+: a hybrid method for predicting
rotein backbone torsion angles from NMR

p ondary chemical shifts. J Am Chem Soc 127:
chemical shifts. J Biomol NMR 44:213–223 14970–14971
25. Shen Y, Bax A (2013) Protein backbone and 32. Kabsch W, Sander C (1983) Dictionary of pro-
sidechain torsion angles predicted from NMR tein secondary structure: pattern recognition of
chemical shifts using artificial neural networks. hydrogen-bonded and geometrical features.
J Biomol NMR 56:227–241 Biopolymers 22:2577–2637
26. Shen Y, Bax A (2010) SPARTA plus: a modest 33. Schwieters CD, Kuszewski JJ, Tjandra N et al
improvement in empirical NMR chemical shift (2003) The Xplor-NIH NMR molecular struc-
prediction by means of an artificial neural net- ture determination package. J Magn Reson
work. J Biomol NMR 48:13–22 160:65–73
27. Jones DT (1999) Protein secondary structure 34. Herrmann T, Guntert P, Wuthrich K (2002)
prediction based on position-specific scoring Protein NMR structure determination with
matrices. J Mol Biol 292:195–202 automated NOE assignment using the new soft-
28. Rost B, Sander C (1993) Prediction of protein ware CANDID and the torsion angle dynamics
secondary structure at better than 70 percent algorithm DYANA. J Mol Biol 319:209–227
accuracy. J Mol Biol 232:584–599 35. Markley JL, Bax A, Arata Y et al (1998)
Recommendations for the presentation of
29. Ramirez BE, Voloshin ON, Camerini-Otero
NMR structures of proteins and nucleic acids
RD et al (2000) Solution structure of DinI
(Reprinted from Pure and Applied Chemistry,
provides insight into its mode of RecA inactiva-
vol 70, pp. 117–142, 1998). J Mol Biol 280:
tion. Protein Sci 9:2161–2169
933–952
30. Maltsev AS, Ying JF, Bax A (2012) Deuterium
36. Wang LY, Eghbalnia HR, Bahrami A et al
isotope shifts for backbone 1H, 15N and 13C
(2005) Linear analysis of carbon-13 chemical
nuclei in intrinsically disordered protein alpha-
shift differences and its application to the
synuclein. J Biomol NMR 54:181–191
detection and correction of errors in referenc-
31. Berjanskii MV, Wishart DS (2005) A simple ing and spin system identifications. J Biomol
method to predict protein flexibility using sec- NMR 32:13–22
Chapter 3
Predicting Bacterial Community Assemblages Using

an Artificial Neural Network Approach
Peter Larsen, Yang Dai, and Frank R. Collart
Abstract
Microbial communities are found in nearly all environments and play a critical role in defining ecosystem
service. Understanding the relationship between these microbial communities and their environment is
essential for prediction of community structure, robustness, and response to ecosystem changes. Microbial
Assemblage Prediction (MAP) describes microbial community structure as an artificial neural network
(ANN) that models the microbial community as functions of environmental parameters and community
intra-microbial interactions. MAP models can be used to predict community assemblages over a wide
range of possible environmental parameters, extrapolate the results of point observations across spatial
scales, and make predictions about how microbial communities may fluctuate as the result of changes in
their environment.
Key words Modeling, Microbial assemblage prediction, Microbial ecology, Systems biology,
Metagenomics
1 Introduction
From boiling lakes [1, 2] to habitats under miles of Antarctic ice

[3], from the upper atmosphere [4, 5] to deep in the planet’s crust
[6–8], bacteria are found not only to thrive but to be critical driv-
ers of Earth’s biological and geochemical systems. Microbial
metabolism contributes to all known biogeochemical cycles and
has both direct and indirect impacts on Earth’s climate. These
organisms have been implicated in mass extinction events [9] and
are predicted to play important roles in future global warming
events [10].
The submitted manuscript has been created by UChicago Argonne, LLC, operator of Argonne National
Laboratory (“Argonne”). Argonne, a US Department of Energy Office of Science laboratory, is operated
under contract no. DE-AC02-06CH11357. The US Government retains for itself, and others acting on
its behalf, a paid-up nonexclusive, irrevocable worldwide license in the said article to reproduce and pre-
pare derivative works, distribute copies to the public, and perform publicly and display publicly, by or on
behalf of the government.
33
34 Peter Larsen et al.
The identity of these key players in Earth’s ecosystems has

until recently remained elusive. The diversity of these communities
is vast, with the majority of microbial species uncharacterized in
the lab and their very ubiquity challenges systematic sampling.
Metagenomic sequencing, the sequencing of bacterial genomic
DNA extracted directly from environmental samples, has opened
these previously intractable ecosystems for analysis [11]. However,
there is a wide divergence between the scale at which metagenomic
sampling can be conducted and the scale of possible mechanisms
by which microbial population may affect ecosystems [12]. This
divergence supports the need for computational tools that leverage
relatively sparse microbial population observations and extrapolate
those biological observations to prediction of population structure
changes over time, space, or environmental conditions [13].
Here, we present a computational method, Microbial
Assemblage Prediction (MAP), for leveraging microbial popula-
tion data using artificial neural networks (ANNs) to model micro-
bial populations, first described by Larsen et al. [14]. MAP uses
an ANN approach to predict a bacterial population structure, to
a given taxonomic resolution, as a function of environmental
parameters. This approach considers not only the impact of envi-
ronmental factors on bacterial taxonomic abundance but also the
interrelationships between the abundance of one taxonomic
group on another. Required data for this approach includes a set
of metagenomic-derived microbial population structure profiles
collected across a range of environmental conditions.
There are two principal steps in the process for construction
of the MAP model (Fig. 1). The first is to generate an environ-
mental interaction network (EIN). The EIN is a directed acyclic
graph that is derived from the statistical dependencies between
environmental parameters and relative abundance of microbial
taxa. The EIN is generated such that the root nodes of the net-
work are always environmental parameters and never microbial
taxa. In the second step, the EIN is used as the underlying topol-
ogy of the MAP ANN. Each node in the ANN has a value that is
defined as a function of the values of its parent nodes with the
root environmental parameter nodes taking values from biologi-
cal observations. In this way, values for all nodes in the ANN are
ultimately mathematical functions of environmental parameter
data. This enables the generation of MAP models of microbial
populations that can be used to predict microbial population
structure across multiple observed or extrapolated environmental
conditions (Fig. 2). Model predictions can be used to direct
future sampling efforts and design hypothesis-generated experi-
ments that interrogate the ecological relationships in microbial
communities.
Predicting Bacterial Community Assemblages 35
Fig. 1 Two steps to generate MAP model. Generation of MAP model requires as
input a set of related microbial population structures and corresponding environ-
mental parameter data. The first step is to generate the environmental interac-
tion network (EIN) (left panel). The EIN is the set of identified statistical
relationships between environmental parameters and microbial taxa abundances
and between microbial taxa abundances, presented as a directed acyclic graph.
In the small, theoretical example above, there are three environmental parame-
ters and four microbial taxa represented in the network. In an EIN, root nodes are
always environmental parameters. The second step of generating a MAP model
(right panel) is fitting the data to equations derived from the EIN. Each value of a
node in the EIN is defined to be a function of the values of its parent nodes. In the
complete MAP model, the relative abundance of every taxa node in the network
is ultimately a function of the value of environmental parameter nodes
2 Materials
2.1 Collected Two kinds of data are required in MAP models: microbial population
Biological structure and environmental parameter data (see Note 1).
Observations While the list of possible environments and techniques for
metagenomic sampling are outside of the scope of this protocol,
methods for metagenomic sample collection and sequencing are
reviewed by Thomas et al. [11]. Raw metagenomic data must be
analyzed to identify the relative abundance of specific taxonomic
groups present in the microbial populations. There are many avail-
able computational tools for this, but “QIIME” for 16 s RNA
sequence data (http://qiime.org/) [15] and “MG-Rast” for shot-
gun metagenomic data (http://metagenomics.anl.gov/) [16] are
examples of excellent and freely available software solutions. The
population data needs to be binned to some level of taxonomy
(e.g., order, class, or genus). The selection of specific taxonomic
level to which data is binned depends greatly on the nature of the
microbial population being modeled but generally should be
Fig. 2 Examples of MAP model applications. MAP models can be used to investigate microbial populations at
a variety of temporal and spatial scales. The included examples drawn from MAP model and data originally
presented in Larsen et al. [14]. (a) In a MAP model extrapolated across time, marine microbial populations are
predicted for a single location over a total of 6 years. For every day in this model, environmental parameters
at a single location in the Western English Channel were inferred from satellite imagery and used as input to a
MAP model. In this figure, gray lines indicate changing relative abundance of individual taxa. Black “X”s indi-
cate the Bray-Curtis similarity score between the predicted and observed population structure at every time
point. (b) In this MAP model extrapolated over space, MAP model generated from data collected at a single
point is used to extrapolate the population structure over a wide geographical area using satellite image mod-
eling data as environmental parameters. In this figure, height of surface is proportional to relative abundance
of the taxa Rickettsiales in the surface waters of the Western English Channel on March 17, 2008. The geo-
graphical region pictured is between 48.5°N, 7.6°W and 50.8°N, 1.5°W. (c) In this figure, MAP model is used
to extrapolate over multiple environmental conditions. Using a MAP model generated using satellite image
data collected between 2007 and 2008, an initial microbial population structure (black bars) was generated
for August 20th, 2008. In the figure, y-axis is percent abundance of total population. X-axis is the set of 24
bacterial taxa in this MAP model. Starting with these observed environmental parameters, environmental
parameters were changed in silico until a “bloom” of Vibrionales was observed in microbial population without
greatly altering the abundance of all other microbial taxa (gray bars, Vibrionales abundance highlighted with
arrow). Strikingly, the conditions required to generate an August Vibrionales bloom in MAP model (i.e., lower
soluble phosphorus, increase in chlorophyll A concentration, and moderate levels of dissolved organic carbon)
are nearly identical to environmental parameters associated with a Vibrionales bloom in August 2003
selected by using the highest taxonomic level for which the majority
of observations in all populations, the observed abundance of a
taxa is nonzero.
Each microbial population structure observation must be
accompanied by a set of environmental parameters. The nature of
this data may vary widely depending on the nature of the environ-
ment being sampled and the scale of the desired MAP model but
may include chemical and physical attributes collected from in situ
observations, data collected from automated sensors deployed in
the environment, or remotely sensed data such as by satellite image.
2.2 Required Some computational tools are required to generate MAP models,
Computational Tools and specific tools are recommended here for use in construction of
MAP models (see Note 2).
For generating EIN, the Bayesian Network Inference tool
“Banjo” [17] is recommended. Banjo is a software application and
framework for structure learning of static and dynamic Bayesian
networks using a score-based structure inference. Banjo is freely
available under a noncommercial use license (http://www.cs.duke.
edu/~amink/software/banjo/).
For generating the equations in the ANN from the topology of
the EIN, Eureqa is recommended. “Eureqa Formulize” is a soft-
ware package available from Nutonian Inc. (http://www.nutonian.
com/), which is freely available for academic research (http://
www.nutonian.com/products/eureqa/). As described by Eureqa
documentation, this analysis tool searches the space of mathemati-
cal equations using symbolic regression to identify a mathematical
model that best fits the data provided. Starting with initial expres-
sions that randomly couple mathematical building blocks, both the
form and parameters of possible equations are generated by recom-
bining previous equations and varying their sub-expressions using
a statistics-based evolutionary algorithm. The resulting models are
ranked using a ratio of equation accuracy and complexity.
3 Methods
There are two steps for construction of the MAP model: genera-
tion of EIN and fitting equations to ANN. Before beginning, the
available data should be divided into training and validation sub-
sets. The training subset is used to construct the MAP model and
then verify the accuracy of the MAP model with the validation
subset.
3.1 Normalize Data Environmental parameter data should be normalized to a uniform

from Biological dynamic range, for example, to arbitrary units ranging from 0 to
Observations 100 (Fig. 3) (see Note 3).
Typical microbial population data is made up of a relatively
small number of species comprising the majority of the population
and a long tail of numerous low-abundance species comprising the
remainder of the population [18]. To best normalize the data dis-
tribution for MAP model, first report taxonomic abundances as a
percentage of total population. This normalized data is then log2
transformed (Fig. 4).
3.2 Generate EIN The following steps are specific to Banjo software (see Note 4).
from Normalized Data Before generating BN network in Banjo, a file of unallowed edges
must be made (“edgesNotAllowed” parameter in Banjo settings file).
This parameter must be set such that environmental parameter nodes
Fig. 3 Example normalization of environmental parameters. This very small,

hypothetical sample dataset contains three environmental parameters and four
observations. Raw data for collected environmental parameters (top panel) can
be in different units and cover very different dynamic ranges. Before using in
MAP models, data must be normalized to uniform ranges in arbitrary units. In this
example (bottom panel), data is normalized such that the lowest observed value
for an environmental parameter is equal to 20 and the highest is equal to 80
have no possible parent nodes. Additional parameters in Banjo, such

as maximum number of parents, searcher, proposer, and run dura-
tion, depend entirely on the nature of data being considered and
must be determined empirically. Investigators should refer to the
Banjo user manual for specific guidelines in selecting search parame-
ters appropriate for a particular dataset.
The output of Banjo is the network that best fits the microbial
population data. The final network is used to generate the equa-
tions in the subsequent step.
3.3 Generate ANN The same, normalized dataset used to generate the EIN is used to
from EIN generate equations for the MAP model ANN. Specifics for running
Eureqa on particular platforms are available from the Nutonian
Eureqa website (http://www.nutonian.com/products/eureqa/).
From the output of Banjo EIN, network topology in the format of
a list of parent nodes for each child node is converted into Eureqa.
The operations “constant,” “add,” “subtract,” “multiply,” “divide,”
and “exponential” are recommended. As Eureqa outputs several
Fig. 4 Example normalization of microbial population data. In this small, hypo-

thetical microbial population dataset, there are four microbial taxa (a–d) and four
observations. As in the top panel, the initial data will be in the format of number
of 16s RNA sequences associated with each taxa. This data is transformed (mid-
dle panel) to representation as percent of total population. Percent total popula-
tion data is log transformed before use in MAP analysis (bottom panel)
possible functions to fit observed data, the equation selected for use
in MAP was selected to be the most parsimonious one as defined by
the following, as many as possible, optimization criteria:
– Smallest equation size.
– Highest correlation with observed data.
– If there is a choice between an equation with an exponential
term and one without, choose the equation without an expo-
nential term.
– If there is an obvious peak or drop in observed relative bacte-
rial abundance, choose the equation that best models that peak
or drop, even if it means selecting an equation with a greater

size or lower overall correlation.
Eureqa’s equation search is “done” when the stability metric is
close to 100 %. The time it takes to achieve stable solutions will
depend on the nature of the data and on the speed of the computer
running Eureqa. It is important to note that even if a node is pre-
dicted to have more than one parent in the EIN, not all of those
parents will necessarily be incorporated into the Eureqa-generated
equations that best fit the observed data.
3.4 Implement The results of the previous step are a set of linked mathematical
MAP Model equations in which the results of some equations will be the terms
in other equations. Taken together, given a single set of normal-
ized environmental parameter data, values for all taxonomic rela-
tive abundances can be calculated with a unique solution.
Implementation of this model can be easily accomplished in a
number of mathematical programming platforms (e.g., R-Project
(http://www.r-project.org/) or MatLab (http://www.mathworks.
com/products/matlab/)) or common spreadsheet applications.
3.5 Validate Once the MAP model has been generated, the model must be vali-
MAP model dated. Similarity between predicted population structures in the
validation subset can be calculated by a variety of means, such as
Pearson’s correlation coefficient, product moment correlation
coefficient, or Bray-Curtis similarity score [19]. The specific simi-
larity metric must be chosen depending on the distribution of the
data used to generate the model. Due to the typical distribution of
microbial population data, in which a few taxa are consistently of
high relative abundance and many taxa are consistently of low rela-
tive abundance [18], it is possible to get deceptively high correla-
tions between predicted and observed population structures so
long as the model correctly identifies the small number of high-
abundance taxa. To correct for this, there are two null models that
can be easily tested. The first null model is to set all taxa’s predicted
relative abundance equal to the average taxa abundance across all
samples. The other null model is to set all taxa abundances equal to
the minimum observed taxa’s abundance. If either of these null
models has a better correlation with biological observations than
the MAP model, it is likely that the MAP model is not identifying
useful relationships in the data (see Note 5).
To assign a level of statistical significance to the MAP model
results, a bootstrap-style approach is recommended (see Note 6).
In a bootstrap-like method, the model data is randomly permuted
a large number of times. The similarity of observed data to per-
muted data is calculated each interaction. A significance of the
model can be reported as:
# iterations in which random permutation similarity  model similarity
Total # of iterations
This permutation approach can be accomplished in a number of

mathematical programming tools or in common spreadsheet
applications.
Identifying a MAP model that both accurately predicts
observed microbial population structures and is likely to provide
useful biological insight is a process that may take several iterations
of model building and testing.
4 Notes
1. As with any biological experiment or computational model,

there needs to be a sufficient number of training data observa-
tions to draw meaningful biological conclusion from MAP
model. The exact number of observations required to achieve
results with statistical significance or biological relevance will
depend entirely on the nature of the data collected and the
complexity of the biological system being studied and must be
empirically determined.
2. Other computational tools that perform the same or similar
functions could also be incorporated, depending on availability
and preference of individual investigators.
3. If using a minimum of 0 and a maximum of 100 for environ-
mental parameter data in MAP model, it may be useful to
normalize observed data to values within that range, for exam-
ple, a minimum of 20 and maximum of 80. This allows the
possibility of modeling environmental parameters slightly out-
side the range of actual observations without running into the
potential computational errors that might be caused by using
input values equal to or less than zero.
4. If an alternate program for BN inference is used, then follow
the documentation for that software.
5. If the resulting MAP model does not yield satisfactory results,
it is necessary to reconsider the parameters used to build the
model. The following are common steps to consider when
troubleshooting a MAP model:
Generation of EIN
– Data discretization: Banjo requires discretized data for
learning networks. It is possible that discretization could
mask important variability in some microbial taxonomic
abundance values. Consider a higher discretization index
or using quantile discretization in Banjo.
– Rare taxa present in data: Some taxa may be undetected in
many samples. Rarely observed taxa tend to be over-fitted
in the network. Consider removing these rare taxa from
the analysis or selecting a lower taxonomic resolution.
– Very sparse networks: Increasing the number of allowed

parents in Banjo requires more computer memory but may
generate networks with higher connectivity.
– Lack of interaction between environmental parameters and
taxa abundance nodes: It is possible that Banjo will not
identify many direct interactions between environmental
parameters and microbial taxa abundances. It may be use-
ful to log transform environmental parameter data or select
an alternate discretization policy in Banjo. If no statistically
significant dependencies between environmental parame-
ters and microbial taxa abundances can be found in Banjo,
it may be necessary to resort to an alternate method, such
as strong correlation coefficients between environmental
parameters and microbial taxa abundance to identify
potential edges in EIN.
Generation of MAP ANN
– Over-fitting of equations by “Eureqa”: It may be that
the training set correlates well with biological observa-
tions, but the validation set correlated poorly. From the
set of Eureqa-proposed equations, reselect an equation
with lower complexity, even if the equation has lower
correlation with training data.
– Unsolvable equations in MAP model: Some combinations
of data may cause Eureqa-generated equations to return
errors, such as attempting to divide by zero. Reselect any
equations for which a division-by-zero error may occur or
consider removing rare taxa from the model for which rela-
tive abundance is sometimes or often equal to zero.
– Biologically meaningless results: It is possible for MAP
model to predict biologically meaningless results such as
negative taxonomic abundances or taxa that comprise
more than 100 % of the total population. To prevent this,
avoid selecting Eureqa equations with very large constants
or exponential terms. Also, it is often helpful to set upper
and lower limits on equation results, setting values less
than zero equal to zero (or an empirically determined very
small number to avoid possible division-by-zero errors)
and establishing an upper limit on possible abundances
(e.g., setting an upper limit equal to a value 20 % larger
than the maximum relative abundance of the most abun-
dant taxa actually observed in the biological dataset).
6. There are many ways to permute the data, and the most appro-
priate way to randomize the data will depend on the nature of
the data and the experimental design. One common approach
to randomization is to randomly reassign predicted abun-
dances within the same taxa and across observations.
References
1. Hugenholtz P, Goebel BM, Pace NR (1998) 10. Bardgett RD, Freeman C, Ostle NJ (2008)
Impact of culture-independent studies on the Microbial contributions to climate change
emerging phylogenetic view of bacterial through carbon cycle feedbacks. ISME J 2(8):
diversity. J Bacteriol 180:4765–4774 805–814
2. Barns SM, Takala SL, Kuske CR (1999) Wide 11. Thomas T, Gilbert J, Meyer F (2012)
distribution and diversity of members of the Metagenomics - a guide from sampling to data
bacterial kingdom Acidobacterium in the analysis. Microb Inform Exp 2(1):3
environment. Appl Environ Microbiol 65: 12. Larsen PE, Gibbons SM, Gilbert JA (2012)
1731–1737 Modeling microbial community structure and
3. Shtarkman YM et al (2013) Subglacial Lake functional diversity across time and space.
Vostok (Antarctica) accretion ice contains a FEMS Microbiol Lett 332(2):91–98
diverse set of sequences from aquatic, marine 13. Larsen P, Hamada Y, Gilbert J (2012) Modeling
and sediment-inhabiting bacteria and eukarya. microbial communities: current, developing,
PLoS One 8(7):e67221 and future technologies for predicting microbial
4. Fierer N et al (2008) Short-term temporal vari- community interaction. J Biotechnol 160(1–2):
ability in airborne bacterial and fungal popula- 17–24
tions. (Translated from eng). Appl Environ 14. Larsen PE, Field D, Gilbert JA (2012)
Microbiol 74(1):200–207 Predicting bacterial community assemblages
5. Bowers RM et al (2009) Characterization of using an artificial neural network approach. Nat
airborne microbial communities at a high- Methods 9(6):621–625
elevation site and their potential to act as 15. Caporaso JG et al (2010) QIIME allows analy-
atmospheric ice nuclei. (Translated from sis of high-throughput community sequencing
eng). Appl Environ Microbiol 75(15): data. Nat Methods 7(5):335–336
5121–5130 16. Meyer F et al (2008) The metagenomics RAST
6. Takai K et al (2001) Archaeal diversity in waters server—a public resource for the automatic
from deep South African gold mines. Appl phylogenetic and functional analysis of metage-
Environ Microbiol 67(12):5750–5760 nomes. BMC Bioinform 9:386
7. Edwards RA et al (2006) Using pyrosequenc- 17. Smith VA et al (2006) Computational infer-
ing to shed light on deep mine microbial ecol- ence of neural information flow networks.
ogy. BMC Genomics 7:57 PLoS Comput Biol 2(11):e161
8. Teske A, Sorensen KB (2008) Uncultured 18. Sogin ML et al (2006) Microbial diversity in
archaea in deep marine subsurface sediments: the deep sea and the underexplored “rare bio-
have we caught them all? ISME J 2(1):3–18 sphere”. Proc Natl Acad Sci U S A 103(32):
9. Baune C, Bottcher ME (2010) Experimental 12115–12120
investigation of sulphur isotope partitioning 19. Mumby PJ, Clarke KR, Harborne AR (1996)
during outgassing of hydrogen sulphide from Weighting species abundance estimates for
diluted aqueous solutions and seawater. Isotopes marine resource assessment. Aquat Conserv
Environ Health Stud 46(4):444–453 6(3):115–120
Chapter 4
A General ANN-Based Multitasking Model for the Discovery

of Potent and Safer Antibacterial Agents
A. Speck-Planche and M.N.D.S. Cordeiro
Abstract
Bacteria have been one of the world’s most dangerous and deadliest pathogens for mankind, nowadays
giving rise to significant public health concerns. Given the prevalence of these microbial pathogens and
their increasing resistance to existing antibiotics, there is a pressing need for new antibacterial drugs.
However, development of a successful drug is a complex, costly, and time-consuming process. Quantitative
Structure-Activity Relationships (QSAR)-based approaches are valuable tools for shortening the time of
lead compound identification but also for focusing and limiting time-costly synthetic activities and in vitro/
vivo evaluations. QSAR-based approaches, supported by powerful statistical techniques such as artificial
neural networks (ANNs), have evolved to the point of integrating dissimilar types of chemical and biologi-
cal data. This chapter reports an overview of the current research and potential applications of QSAR
modeling tools toward the rational design of more efficient antibacterial agents. Particular emphasis is
given to the setup of multitasking models along with ANNs aimed at jointly predicting different antibacte-
rial activities and safety profiles of drugs/chemicals under diverse experimental conditions.
Key words Antibacterial activity, Drug resistance, QSAR, Topological indices, Ontology, Moving
average approach, Artificial neural networks, mtk-QSBER models
1 Introduction
For centuries, mankind has been particularly affected by microbial

diseases, those caused by bacteria being among the most danger-
ous and lethal. Even though antibiotics have saved millions of lives
and alleviated the suffering of people for many years [1], drug-
resistant, disease-causing bacteria have emerged and spread so
much in recent decades [2–5] that they have created a global pub-
lic health problem. This situation is so serious that nowadays bac-
teria cause high mortality rates not only in communities but also in
healthcare environments and facilities [6]. Given the prevalence of
the major gram-positive pathogens, especially those belonging to
the genera Staphylococcus, Streptococcus, and Enterococcus, transfer-
ability of resistance genes is of special concern. This is exemplified
45
46 A. Speck-Planche and M.N.D.S. Cordeiro
by the ability of enterococci to interact with methicillin-resistant

S. aureus (MRSA), transferring to the latter vancomycin-resistant
genes [7]. Thus, MRSA can become resistant to vancomycin,
which is the unique antibacterial chemotherapeutic alternative. On
the other hand, gram-negative pathogens such as Escherichia coli
and Pseudomonas aeruginosa exhibit intrinsic resistance to almost
any antibiotic [8]. The most deadly bacterial disease has been
tuberculosis (TB), which is prone also to multidrug resistance.
Indeed, more than 9 million new TB cases and about 1.7 million
deaths were reported in 2009 [9], while a recent report confirmed
that in 2010, 5.7 million official cases of TB were notified, with 8.8
million incidents worldwide (equivalent to 128 cases per 100,000
population) [10].
Diseases and infections caused by bacteria are very complex
and depend on multiple pathogenic and epidemiological factors
[6, 8]. For this reason, the search for new antibacterial drugs is
very urgent. This battle against bacterial diseases/infections will
depend on the efficiency of chemotherapies and have a profound
influence on human health.
High-throughput screening technologies [11] along with
combinatorial chemistry [12] were expected to solve the drug dis-
covery problem by a massive parallelization of the process. In prac-
tice, however, while the number of identified hits can substantially
be increased using these methods, no corresponding growth in the
number of drugs entering the market has been observed [13]. This
fact has progressively led to a reconsideration and rationalization of
the drug discovery process, in which chemoinformatics methods,
led by Quantitative Structure-Activity Relationships (QSAR) tools
[14], have gained a role of tremendous importance [15–17]. To be
clearly effective, these methods should aim at targeting both phar-
macological profiles and desirable ADMET (absorption, distribu-
tion, metabolism, elimination, toxicity) properties, as the latter are
the major causes of non-approval of drugs [18–20]. Particularly, in
recent years, promising multi-target (mt-QSAR) models have been
reported, revolutionizing concepts regarding QSAR paradigm pre-
diction [21–26]. These mt-QSAR models have been applied to
simultaneously predict several features of biological activity by con-
sidering different biological targets such as biomolecules, cell lines,
tissues, and organisms. The aforementioned mt-QSAR models
have evolved to the point that nowadays it is possible to predict
multiple biological profiles by considering different measures of
the effects, many biological targets, and diverse levels of curation
of the experimental information [27–32]. These new models,
known as multitasking Quantitative Structure-Biological Effect
Relationships (mtk-QSBER), have often employed classification
techniques such as linear discriminant analysis (LDA) [33]. But
sometimes, due to the complexity of the data and a lack of accuracy
in the experiments, linear classification tools do not lead to models
ANN-Based Mtk-Model for Discovery of Antibacterial Agents 47
with enough statistical quality and predictivity. In such cases,

artificial neural networks (ANNs) can instead be applied to deal
with nonlinear behavior and usually do enhance the performance
of the predictions [22, 29]. This chapter is devoted to an overview
of the setup and use of mtk-QSBER models based on ANNs.
Overall, specific insights from our perspective and personal experi-
ence will be provided.
2 General Procedure for the Setup of QSAR Models
Generally speaking, QSAR approaches seek to uncover possible

relationships between chemical structures (molecular descriptors)
and one or more endpoints of interest (e.g., biological activity, tox-
icity, or pharmacokinetic profiles) [14]. Several steps should be fol-
lowed for establishing the QSAR models (Fig. 1), and these, if well
devised, may help yielding more accurate results. Firstly, the data
must be retrieved typically from a public source and subsequently
curated in, for example, a spreadsheet application like Excel.
Secondly, molecular descriptors are calculated from the molecular
structures; finally, by applying a statistical modeling approach
(either linear or nonlinear), the best QSAR model can be obtained.
2.1 Databases Databases are typically organized as public sources and contain
and Handling large collections of data describing the most relevant aspects of
of the Retrieved Data phenomena and/or experiments [34, 35]. A well-known database
is ChEMBL [34]—an online free and dynamic source available at
Fig. 1 Necessary steps required for building a QSAR model

https://www.ebi.ac.uk/chembl/, which contains more than 12

million assay outcomes. These biological assays derive from apply-
ing more than 1.3 million drugs/chemicals against at least one out
of more than 9,300 biological targets (proteins, cell lines, microor-
ganisms, mammals), and it is possible to obtain specific informa-
tion about the diverse strains or breeds in the case of non-molecular
biological entities. There is also information related to the type of
the assay, i.e., if a test was carried out to measure binding (B), func-
tional (F), or ADMET (A) properties. All data can be easily down-
loaded into Excel-like files. Most importantly, the ChEMBL
database provides good transparency concerning the accuracy
and/or reliability of the experimental information. For that, two
columns of ChEMBL are of particular interest, namely, (1) a
“CONFIDENCE SCORE” that contains values ranging from 4 to
9 for biomacromolecular targets, the highest number being given
to a very accurately determined experiment, and (2) a “CURATED
BY” that shows how much the experimental information has been
verified, depending on the availability of data in the literature and
certain details associated to the assay protocols, with three catego-
ries: autocuration (the lowest reliability), intermediate, and expert
(highest reliability). In addition, in this database, each compound
has one ChEMBL identifier and SMILES (Simplified Molecular
Input Line Entry Specification, i.e., the 2D chemical structure) code.
Therefore, ChEMBL is an interesting and potentially valuable
resource for tackling drug discovery. In fact, it has been stated that
mining ChEMBL is an excellent alternative to generate models for
virtual screening of large libraries of drugs and compounds [36],
allowing a greater coverage of the chemical space. With these con-
siderations in mind, ChEMBL should be the database of first choice.
A further very important database is DrugBank, which is avail-
able at http://www.drugbank.ca/. This is a unique free source
chemo-bioinformatics database that combines specific drug infor-
mation (chemical structure, pharmacological, and therapeutic
data) with details from the respective target(s) (sequences, struc-
tures, and pathways) [35]. DrugBank comprises 6,825 drug entries
including 1,541 small-molecule drugs approved by the FDA (Food
and Drug Administration) and other molecular entities such as 86
nutraceuticals, 5,082 experimental chemicals, and 150 FDA-
approved biotech (protein/peptide) drugs. DrugBank is a suitable
source of raw data for deep studies of drug-target and drug-drug
interactions. QSAR modeling using this database can facilitate the
discovery of new targets for antibacterial drugs. At the same time,
studies focused on drug-drug interactions can be performed in
advanced stages of the drug discovery process to analyze, e.g., pos-
sible chemicals with which an antibacterial drug should not be
administrated. In addition, ChEMBL and DrugBank offer the
possibility of downloading substantial volumes of data in the form
of tabular-based files which can be easily opened, modified, and
filtered/curated using Excel.
2.2 Molecular Molecular descriptors are essential hot spots of any QSAR study.
Descriptors and Their As pointed up nicely some years ago by Todeschini and Consonni
Applicability in QSAR [37]: “The molecular descriptor is the final result of a logic and
Approaches mathematical procedure which transforms chemical information
encoded within a symbolic representation of a molecule into a
useful number or the result of some standardised experiment.”
The use of certain types of molecular descriptors depends on the
kind of QSAR approach that is going to be employed. Molecular
descriptors used in classical QSAR approaches may be experimen-
tally determined (e.g., physicochemical properties like the
Hammett σ constant, refractivity, etc.) via Hansch analysis or
through the use of simple indicator descriptors (R-group substitu-
tions of the parent core), Free-Wilson analysis [38]. 3D-QSAR
approaches have emerged as a natural extension to the classical
QSAR approaches pioneered by Hansch and Free-Wilson and are
based on correlating the activity with non-covalent molecular
interaction fields [38]. These approaches require mandatorily 3D
structures, e.g., based on protein crystallography or molecule
superimposition, and assume that all molecules interact with the
bio-target in the same manner, i.e., with an identical binding mode.
In 3D-QSAR approaches, pair potentials (e.g., Lennard-Jones) are
calculated for the whole molecule to model the effects of shape
and size (steric fields), the presence of hydrophobic/hydrophilic
regions, and the behavior of the electronic density in different
parts of the molecule (electrostatic fields). CoMFA (Comparative
Molecular Field Analysis) and CoMSIA (Comparative Molecular
Similarity Index Analysis) [39] are the most applied 3D-QSAR
approaches, in particular, to support docking calculations. Over
time, 3D-QSAR approaches have evolved to improved variants
such as 4D-QSAR, additionally including ensemble of ligand con-
formations and protonation states in 3D-QSAR [40]; 5D-QSAR,
adding induced-fit effects of the adaptation of the receptor-
binding pocket to the ligand in 4D-QSAR [41]; and 6D-QSAR
approaches, further incorporating different solvation models in
5D-QSAR [42].
As an alternative to 3D-QSAR approaches, the structures of
the molecules can be optimized without any superimposition of
alignment. Toward that end, calculations based on, e.g., density
functional theory (DFT) and molecular dynamics (MD) are car-
ried out [43], and both local (charge, polarizability HOMO-/
LUMO-based parameters) and global descriptors (different ener-
gies, volumes, and areas) can be determined and used to obtain
correlations with biological activities. Further, highly accurate
quantum-mechanics calculations allow the study of more specific
interactions such as the strength of hydrogen interactions, stability
of coordination bonds, etc.
Several works using 3D-QSAR approaches or QM calculations
have been reported in the field of antimicrobial research [44–48].
But from our point of view, QSAR research should be focused on

developing models best suited for the analysis of a large number of
diverse compounds and virtual screening of databases. For that
purpose, 3D-QSARs and related approaches have several impor-
tant drawbacks, namely:
–– Although improved approaches such as 4D-, 5D-, and
6D-QSARs have emerged, uncertainties regarding the selec-
tion of the active conformations and the binding mode of
ligands remain.
–– 3D-QSARs are usually used to study the inhibitory activity of
compounds related to in vitro data. For the successful applica-
tion of these approaches to in vivo data, there must be a thor-
ough understanding of the binding mechanisms of all
underlying ligands.
–– For 3D-QSAR approaches, optimization of the 3D structures
of ligands can be particularly time consuming. Moreover, the
lowest energy conformation of any ligand is usually considered
to be the bioactive conformation and responsible for exerting
the binding effects.
–– In the case of QM-based models, the computational cost and
required time to perform calculations can be very high,
depending on the complexity of the molecules to be modeled,
the QM method to be applied (ranging from molecular
mechanics to semiempirical and DFT methods, including
higher-level QM methods), and the availability of computa-
tional resources.
–– If QM calculations are performed without alignment, there
will be more uncertainty than in 3D-QSAR approaches for the
selection of the active ligand conformation.
These handicaps can somehow be solved if 2D topological
descriptors are applied [37]. More than 100 families of topologi-
cal descriptors have been described in the literature [49]. They are
among the most useful sets of molecular descriptors, as corrobo-
rated by the large number of QSAR applications reported to date
[50–60]. However, these descriptors have been criticized for a
lack of clarity regarding their physicochemical meaning and struc-
tural interpretation [38]. Definitely, these descriptors are unable
to provide a complete understanding of the 3D structural aspects
of molecules, which can be considered as a possible disadvantage
if the analysis of the binding mode of a ligand with its bio-target
is needed. But as proven by Estrada, although topological descrip-
tors are derived from 2D representations of the molecular
structures, connectivity indices, for example, encode information
related to the molecular accessibility for the surrounding medium.
For instance, first- and second-order connectivity indices repre-
sent molecular accessibility areas and volumes, respectively, while
higher-order connectivity indices represent hyper-volumes [61].

The same author has also demonstrated that 2D topological
descriptors can account for “pure” 3D structural parameters, such
as the dihedral angle between phenyl rings in alkylbiphenyls [60].
At this point, it is easy to conclude that topological descrip-
tors do not depend on the conformation and thus do not need
the superimposition rules and alignments used in 3D-QSAR
approaches. Further, topological descriptors can be used in QSAR
studies involving both in vitro and in vivo assay data. When these
descriptors are used to generate QSAR models based on statistical
classification techniques like LDA, the databases can be formed by
compounds belonging to dissimilar chemical families, because the
mechanisms of action are not so important. These descriptors will
afford the structural patterns to successfully separate the dataset of
compounds into groups having the observed biological activity or
not [21–32, 54, 59]. Topological descriptors have exhibited a
great applicability for virtual screening of antibacterial agents [62–
66]. Finally, extremely interesting kinds of topological descriptors
are the graph-based spectral moments, designed according to the
TOPS-MODE (TOPological Substructural MOlecular DEsign)
approach [67–69]. The greatest advantage of the TOPS-MODE
approach, over other traditional QSAR methods, stems from its
substructural nature. This means that one can transform the result-
ing QSAR model into a bond additive scheme and thus describe
the endpoint activity as a sum of bond contributions related to
different structural fragments of the molecules. Moreover, one can
detect the fragments on a given molecule that contribute positively
or negatively (by summing up bond contributions) to the underly-
ing activity [51].
Many programs have been built for the calculation of molecu-
lar descriptors. One of the most widely applied is the MODESLAB
software [70] developed by Estrada and Gutierrez for Windows.
This software allows the calculation of physicochemical properties
(e.g., polar surface area, van der Waals radii, etc.) and classical
topological descriptors like atom and bond connectivity indices, as
well as other indices [49]. The program also calculates topographi-
cal descriptors resulting from mixing topological indices with QM
calculation features. But the most powerful set of descriptors
implemented in MODESLAB is the spectral moments. The
MODESLAB software is freely available through the webpage
http://www.estradalab.org/links/index.html.
Another computer program which has attracted the attention
of QSAR practitioners is PaDEL-Descriptor, devised by Wei [71].
The current version of this Java-based software calculates 905
descriptors (770 1D and 2D descriptors, as well as 135 3D descrip-
tors) and ten types of fingerprints. These are calculated principally
using the well-known Chemistry Development Kit. PaDEL-
Descriptor is a free open-source software, which can work on all
major platforms (Windows, Linux, MacOS), supporting more

than 90 dissimilar file formats to encode chemical information of
the molecules. This software can be downloaded from http://
padel.nus.edu.sg/software/padeldescriptor/.
But perhaps the best known program for the calculation of
molecular descriptors is DRAGON [72]. In the latest version of
this program, 4,885 molecular descriptors can be calculated for
small, medium, and even larger molecules. Furthermore, after
computing the molecular descriptors for a given dataset, correla-
tions between variables can be analyzed by applying the same soft-
ware in order to exclude redundancies. At the same time, almost
any version of DRAGON allows one to carry out principal compo-
nent analysis (PCA) for feature selection, in order to identify the
variables which best explain the variability of the data. The
DRAGON software is available at http://www.talete.mi.it/prod-
ucts/dragon_description.htm, covering different academic and
commercial licenses.
2.3 Modeling The last stage pertains to the creation of a statistically significant
Techniques QSAR model, where the molecular descriptors serve as the inde-
pendent variables and the desired endpoint response(s) as the
dependent variable(s). For that purpose, there are many different
modeling techniques available in a variety of software packages.
This section will just focus on two of the most commonly applied
statistical techniques to generate QSAR models. The first group
embraces regression approaches like partial least squares (PLS)
[33], fundamentally used in 3D-QSAR studies, and the traditional
multiple linear regression (MLR) [38]. The second group com-
prises classification approaches such as LDA and ANNs.
A very important aspect should be highlighted here, whenever
one is retrieving a large volume of data to develop the QSAR
model. Usually databases are compilations of biological and chemi-
cal data reported in the literature, which have been determined
within different laboratories by diverse workers and using several
types of assay protocols, consequently making it difficult to esti-
mate the associated experimental errors. As regression techniques
try to predict the real response values of the compounds, it is then
evident that if one is working with a large and heterogeneous data-
set of chemicals, it would be very difficult to find a good predictive
QSAR model using such techniques because it is almost impossible
to control the uncertainty of the data. The aim of classification
techniques is to generate lines, planes, hyperplanes, and/or surfaces
able to separate compounds into different groups (e.g., active/
inactive). So, one has only to preestablish the cutoff value(s) of the
endpoint response(s) under study to then categorize the compounds
accordingly. Our personal experience, and looking at diverse pub-
lished studies in the past, led us to conclude that when classifica-
tion techniques are employed even for the prediction of diverse
endpoints, there is a greater possibility of obtaining high-quality

QSAR models [21–32, 54, 59]. This reflects the fact that when
classification techniques are used, there is no need to predict the
exact value(s) of the endpoint(s) under study, so problems related
to the uncertainty of the data are significantly reduced.
However, as previously commented, sometimes simple model-
ing techniques like LDA are unable to cope with the complexity of
the data, and thus ANNs are required for gathering a deeper
knowledge about a particular target phenomenon [22, 29, 31].
Any ANN is an effort to emulate biological intelligent systems
(human brain) by simulating their structure and/or functional
aspects [73]. ANNs entail simple processing units (neurons) that
are linked by weighted connections (synapses), with the output
(axon) signal transmitted through the neuron’s outgoing connec-
tion. To the best of our knowledge, three types of ANNs have been
widely used (Fig. 2) in drug discovery [22, 29, 31], namely, linear
neural networks (LNN), multilayer perceptron (MLP), and radial
basis function (RBF) networks. LNN consists of a first entry layer,
comprising the input nodes directly associated to the molecular
descriptors, and a final output layer—the predicted response. MLP
and RBF include additionally a second layer, known as the hidden
Fig. 2 Different architectures of artificial neural networks

layer, consisting of an array of neurons that receive, transform, and

transmit the incoming signals from the first layer. The functions
applied on the hidden layer are very important—the so-called acti-
vation functions [74], because they model the signals coming from
the input layer, as well as the signal from the hidden layer to the
output layer. LNN models are very similar to those which can be
obtained through LDA, but the algorithms to optimize them are
different. LNNs are conceived to establish direct relationships
between the inputs (molecular descriptors) and the output, i.e.,
the endpoint activity usually expressed as a binary variable. MPLs
use linear combinations of weights from the input layer to the hid-
den layer, and the signal from the hidden layer to the output is
modeled by a nonlinear activation function (usually a sigmoid
function). In RBF networks, nonlinear activation functions (usually
Gaussian functions) are applied in the first step, i.e., from the input
layer to the hidden layer. Following on, a linear combination of
such functions is applied to generate the output layer. Regarding
the nomenclature, as shown in Fig. 2, for instance, the second
ANN has the form MLP 11:11-8-1:1, meaning that this is an MLP
(multilayer perceptron) containing 11 variables (descriptors) which
were used as entries to generate 11 input nodes in the first layer, 8
nodes in the second layer, and one node in the third (output) layer
from which one response variable (endpoint activity) is to be pre-
dicted. Similarly, the first and the last ANNs shown have profiles
LNN 11:11-1:1 and RBF 11:11-305-1:1, respectively. Despite the
increasing use of ANNs in different areas [22, 29, 31], in the last
15 years, few papers have described the use of ANN-based models
to predict antibacterial profiles of compounds [48, 75–77]. In
these works, some models were derived with the aim of predicting
a reduced series of analogue compounds [48, 75, 77] or heteroge-
neous but using relatively medium-sized datasets of chemicals
[76]. This further illustrates the need to use ANNs with the aim of
improving the discovery of desirable antibacterial drugs.
Of the software available for carrying out these modeling
techniques, we will mention only two programs because of their
applicability, user-friendly nature, and availability. One of the pro-
grams widely used in different fields of modern science is
STATISTICA [73]. This is a software package developed by
StatSoft, which affords data analysis and statistics, as well as data
visualization and data mining procedures. The latest version of
STATISTICA can handle large amounts of data and is able to
tackle various file formats. This software can also generate codes
for posterior programming after the QSAR model is built. There
are different classification techniques implemented in STATISTICA
such as LDA, ANNs, support vector machines (SVM), classifica-
tion trees (CT), random forest (RF), and many others [73].
Several types of commercial and academic licenses are available for
different versions of STATISTICA at http://www.statsoft.com/.
In the past, STATISTICA has often been applied to construct

QSAR models aimed at predicting antibacterial activity of hetero-
geneous series of compounds [62–64, 66].
Another freely available program to perform statistical analysis
is Weka [78]. This Java program can be viewed as a collection of
machine learning algorithms for data mining purposes. The algo-
rithms can either be applied directly to a dataset or accessed from the
user’s own Java code. Weka contains tools for data preprocessing,
classification, regression, clustering, association rules, and visual-
ization and a powerful tool for feature selection [78]. It is also
well suited for developing new machine learning schemes. Weka
can be downloaded from http://www.cs.waikato.ac.nz/ml/weka/
downloading.html.
3 An mtk-QSBER Model Combined with ANN for Virtual Screening

of Antibacterial Agents
Despite the wide applicability of alignment-free QSAR models

based on topological descriptors, an unfavorable aspect of such
models is that the antibacterial activity is predicted nonspecifically,
meaning that any compound can be predicted to have antibacterial
activity but without information regarding the bacteria against
which it may be active [76]. What is more, antibacterial activity
data should be integrated with other relevant properties such as
those related to ADMET profiles to effectively discover reliable
antibacterial agents. Of particular interest is that the developed
models can potentially guide the screening and discovery of effec-
tive antibacterial agents. The next subsections will describe the
series of steps involved in building up mtk-QSBER models based
on ANNs, which in principle overcome these limitations. Details of
the diverse steps for the development of mtk-QSBER-ANN mod-
els are depicted in the flowchart of Fig. 3. As a typical example
application, we will describe recent work that reported an mtk-
QSBER-ANN model whose role was to simultaneously predict the
anti-enterococci activity and toxicity in laboratory animals of a set
of compounds [27].
3.1 Curation In the work referred to above [27], 13,073 endpoints belonging
of the ChEMBL to more than 9,000 chemicals/drugs were retrieved from ChEMBL
Database in an Excel-compatible file format [34]. Chemicals for which
data were incomplete were eliminated, as were duplicates (i.e., the
same chemicals assayed under the same experimental conditions),
leading to a final dataset comprising 8,560 chemicals/drugs and
to 10,918 statistical cases, considering the various experimental
conditions (cj). The experimental conditions cj cover an ontology
of the form cj ⇒ (me, bt, ai, lc), where me describes the different
measures of biological effects (anti-enterococci activity or toxicity),
Fig. 3 General steps involved in the development of an mtk-QSBER model
and bt denotes the dissimilar biological targets such as enterococci

(including their respective strains), human immune system cells
(lymphocytes), and laboratory animals like mice (Mus musculus)
and rats (Rattus norvegicus). Additionally, ai refers to the specific
assay information, meaning that it provides details whether the
assay focused on the assessment of functional (F) or ADMET pro-
files (A), and lc describes the degree of curation or verification of
the experimental information of a particular assay. Therefore, the
elements me, bt, ai, and lc embody factors that may be modified to
obtain a specific experimental condition.
One should point out also here that the 10,918 cases were the
results of using 8,560 chemicals/drugs to assess one out of 18
measures of biological effects, against at least one out of 131 bio-
logical targets, by considering at least one out of two labels of assay
information, with at least one out of three categories of curation of
the experimental information. An important detail to take into
account is that the measures of antibacterial activity (e.g., the mini-
mum inhibitory concentration, MIC) were often expressed in dif-
ferent units, such as nM or μg/mL, and something similar applies
also to the toxicity measures (e.g., the median lethal dose, LD50).
It is clearly essential that all data for a defined measure be in the same
units; for instance, antibacterial activity should be converted to nM

or a multiple such as μM. All these conversions can be easily made
and are necessary if one is to correctly compare all of the com-
pounds’ biological effects (anti-enterococci potency and toxico-
logical profiles). Further, each of the 10,918 cases was assigned to
one out of two possible groups related with the biological effect
[BEi(cj)] of a defined compound i in a specific experimental condi-
tion cj. That is, a chemical/drug was assigned to the group of posi-
tives [BEi(cj) = + 1] when its value of biological effect fulfilled
certain requirements within a preestablished cutoff [27]; other-
wise, the compound was considered as negative [BEi(cj) = − 1].
3.2 Moving Average A useful predictive mtk-QSBER model clearly depends on the
Approach molecular structure of the compound and on the experimental con-
ditions/ontology cj under which the compound was assayed. But
if the original descriptors are calculated as previously described (see
Subheading 2.2), one will not be able to differentiate the biological
effects for a given molecule when different elements of the ontol-
ogy cj are modified. For this reason, a new set of molecular descrip-
tors was computed by applying the moving average approach [73]:

( ) ( )
Di c j = Di − Di c j
avg

(1)
In Eq. 1, Di is the molecular descriptor of the compound i. The

descriptor Di(cj)avg characterizes a defined set of nj compounds
tested under the same experimental conditions cj. This descriptor is
calculated as the average of the Di values for the compounds in
subset nj that were considered as positive (desirable) cases
[BEi(cj) = + 1] for the same element of cj. It should be clarified here
that, though we have resorted to an arithmetic mean for comput-
ing the Di(cj)avg descriptors, which in fact is one of the best known
measures of the central tendency of a list of data, other measures
like geometric, harmonic means or standard scores could have
been used instead or in addition.
To sum up, ΔDi(cj) are clearly very important descriptors
because they encode information based on the chemical structure
and the characteristics of cj. That is, each ΔDi(cj) represents, in
structural terms, how much a compound deviates from the group
of molecules which were classified as positive.
3.3 Setup Firstly, the dataset under study (10,918 cases) was randomly
of the mtk-QSBER divided into two series: training and prediction sets. The training
Model set contained 8,298 cases, 4,217 assigned as positive and 4,081 as
negative. The prediction or external set included 2,620 cases,
1,353 positive and 1,267 negative cases.
Then, the original molecular descriptors Di were calculated—
i.e., in this case, MODESLAB spectral moments [70] and those of
the type ΔDi(cj) determined by applying Eq. 1. The total final
number of ΔDi(cj) descriptors can be found by multiplying the

number of original descriptors by the number of elements of the
ontology cj. In this work, we have thus started with 120 ΔDi(cj)
descriptors. As usually happens, this is a large number of molecular
descriptors, and so the next task encompasses selecting a proper
subset of descriptors to then build the mtk-QSBER model. In this
work, we have resorted to program Weka [79], which contains a
series of powerful algorithms for variable selection. Notice that the
final set of descriptors should yield the best, or at least a very good,
discrimination between positive and negative groups.
Two algorithms are usually followed for selecting the variable
subset using Weka. In the first, known as the filter algorithm (attri-
bute evaluator), an independent assessment based on the general
characteristics of the data is performed. The second algorithm is
used in combination with the first one, being focused on the subset
evaluation using a defined machine learning algorithm. The latter
is called the wrapper (search) algorithm, because a machine learn-
ing algorithm is wrapped into the selection procedure. More details
about the filter and wrapper algorithms implemented in Weka can
be found in Witten et al. [80]. We suggest using at least a mini-
mum of two combinations of filter and wrapper algorithms, “run-
ning” the data with such combinations at least five times. In so
doing, one is able to easily select the variables common to all such
runs, which embrace the most important attributes. In the case
under study [27], combinations of the attribute evaluator
CFsSubsetEval and the wrapper algorithms called BestFirst and
GeneticSearch were employed. Another possibility might be to use
the forward stepwise (FS) technique as the variable selection strat-
egy, generally along with LDA. After that, the chosen descriptors
can be applied as inputs to derive the best mtk-QSBER model
based on ANNs. Our previous analyses and results have shown us
that when LDA is used with FS, there is a high probability of then
obtaining ANNs of the type RBF. In our opinion, several tech-
niques of variable selection should be combined to provide a solid
background on which molecular descriptors do have more influ-
ence in the final model.
To find the best mtk-QSBER model based on ANNs, LNN,
MLP, and RBF profiles were considered. A fourth ANN called a
probabilistic neural network (PNN) was also analyzed. The soft-
ware STATISTICA was used to carry out this procedure [81]. This
program has a package called “Intelligent Problem Solver,” which
contains a huge number of tools, and one of them allows the evalu-
ation of the most important variables, by performing a task known
as sensitivity analysis. That is, with the latter, only the variables
with sensitivity values higher than one are chosen to enter in the
final model. “Intelligent Problem Solver” should be run at least
five times to check enough different architectures and quality of
the resulting ANNs. Another important aspect is to ensure that at
least one descriptor belonging to each element of the ontology cj is

in the final mtk-QSBER model. Moreover, possible correlation
between descriptors should be analyzed because, if some of the
selected descriptors are redundant, they should be discarded, since
the stability of the model as well as its ability to make future predic-
tions might be affected.
In this work, the best model for the simultaneous prediction of
anti-enterococci activity and toxicity in laboratory animals was
found to be an ANN with an RBF 5:5-767-1:1 profile [27]. As can
be judged from the results in Table 1, the RBF ANN had a far bet-
ter accuracy and predictive power than the LNN—overall accuracy
of around 59 %, two MLPs—overall classification ranging in the
interval 72–77 %, and PNN—accuracy of around 61 %.
The quality and predictive power of the final mtk-QSBER
model were also assessed by checking overall and group-specific
performance measures on the training and prediction sets,
respectively (Table 2). These included the sensitivity, percentage
Table 1
Comparative analysis of the different ANNs exploited in this study
Training set Prediction set
ANNa Sensitivity (%) Specificity (%) Sensitivity (%) Specificity (%)

LNN 5:5-1:1 58.88 58.61 58.17 61.01
MLP 5:5-8-1:1 72.11 72.38 72.28 73.72
MLP 5:5-7-10-1:1 77.35 77.82 77.01 78.06
RBF 5:5-767-1:1 92.98 93.58 90.76 92.11
PNN 5:5-8298-2-2:1 95.04 27.96 94.38 28.02
The first MLP is a three-layer perceptron (TLP), while the second MLP is a four-layer perceptron (FLP)
a
Table 2
Quality and classification performance of the mtk-QSBER model
Classificationa Training set Prediction set

Sensitivity/TP (%) 92.98 90.76
Specificity/TN (%) 93.58 92.11
Accuracy (%) 93.28 91.41
MCC 0.866 0.828
AUROC 0.981 0.965
a
TP, TN, MCC, and AUROC stand for the true positive, true negative, the Matthews
correlation coefficient, and for the area under the ROC curves, respectively
of actual positives that are correctly identified as such or true

positive rate; specificity, true negative rate; accuracy, percentage
of correct classification for all cases; the Matthews correlation coef-
ficient (MCC) [82]; and the area under the receiver-operating
characteristic (ROC) curve [83]. The latter allows one to confirm
that the model is not a random classifier, that is, a model that only
correctly predicts half of the cases (with an area under the ROC
curve of 0.5).
As can be seen in Table 2, the areas under the ROC curves for
the mtk-QSBER model in training and predictions sets were
0.981 and 0.965, respectively, indicating that the model is not a
random classifier. Also, the attained MCC values further corrob-
orate the very good quality and performance of the proposed
mtk-QSBER model, since for both the training and prediction
sets they are near to one. MCC will return values between −1 and
+1, with +1 representing a perfect prediction, 0 a random predic-
tion, and −1, a total disagreement between observed and pre-
dicted biological effects.
3.4 Virtual Screening Many scientists argue that QSAR models are not entirely validated
of Antibacterial as long as no novel compounds are synthesized and biologically
Agents tested. Yet a QSAR model is feasibly validated if the external cases
(not used to build the model) are correctly predicted. That is the
major reason for splitting the datasets into training and prediction
series. The critical point here, however, is the chemical space that
a derived model can cover. The greater the number of different
families of compounds used to build the model, the greater will be
the chemical space in which the model can be used prospectively.
Besides, a well-validated QSAR model should also show promis-
ing results when applied on the virtual screening of chemicals/
drugs. The best way to reveal the promising applications of the
present model was to predict the multiple biological effects of the
antibacterial agent known as BC-3781 [84]. This investigational
drug was reported to exhibit high inhibitory activity against dif-
ferent strains belonging to the genus Enterococcus. By applying
our model, BC-3781 was predicted as a possible antibacterial
agent against different enterococci using multiple experimental
conditions, in good agreement with the experimental evidence.
No toxicological data could be obtained from the literature for
this investigational drug, but the mtk-QSBER model led us to
infer that BC-3781 should not be potentially harmful for labora-
tory animals and, in principle, toxicologically safe for humans as
well. Notice here that although there is still no experimental data
about the toxicity of BC-3781, our toxicity predictions explain
why this compound is already undergoing phase II clinical trials
and that no significant toxic/adverse effects have been reported
so far for humans.
4 Conclusions and Future Perspectives
Nowadays, modern societies are aware of the devastating power of

bacterial diseases and infections. The process of creating innovative
antibacterial chemotherapies can be accelerated by using chemoin-
formatics approaches such as QSAR tools supported by powerful
statistical techniques like ANNs. In this chapter, we have presented
a general overview of the evolution of QSAR models in antibacte-
rial drug discovery. Particular attention was paid to test the ability
of a recent ANN-based mtk-QSBER model in the discovery and
virtual screening of antibacterial agents. It was shown as well how
these types of models are able to participate in dissimilar stages of
drug discovery. In our opinion, future perspectives regarding the
use of mtk-QSBER models combined with ANNs may be focused
on extending them by forward-integrating data of other biological
assays against relevant bacteria such as staphylococci, gram-negative
pathogens, bacteria causing diseases like pneumonia, or those
involved in the appearance and development of nosocomial infec-
tions. As a final point, one should remark here that ANNs can be
viewed as graphs of interconnected nodes. For this reason, the use
of complex network theory may well be useful to analyze more
deeply the topology of these machine learning techniques. Perhaps,
new horizons could be opened and applied to the field of antibac-
terial research.
Acknowledgments
A. Speck-Planche acknowledges the Portuguese Fundação para a

Ciência e a Tecnologia (FCT) and the European Social Fund for
financial support (grant SFRH/BD/77690/2011). This work has
been further supported by FCT through grant N° Pest-C/EQB/
LA0006/2011.
References
1. Grayson ML, Crowe SM et al (eds) (2010) potential impact on empirical therapy. Clin
Kucers’ the use of antibiotics. A clinical review Microbiol Infect 14(Suppl 6):2–8
of antibacterial, antifungal, antiparasitic, and 5. Gonzales R, Corbett KK et al (2008) Drug
antiviral drugs, 6th edn. CRC Press, Taylor & resistant infections in poor countries: a shrink-
Francis Group, LLC, Boca Raton, FL ing window of opportunity. BMJ 336:
2. Shatalin K, Shatalina E et al (2011) H2S: a uni- 948–949
versal defense against antibiotics in bacteria. 6. Lautenbach E, Abrutyn E (2009) Healthcare-
Science 334:986–990 acquired bacterial infections. In: Brachman PS,
3. Cordero OX, Wildschutte H et al (2012) Abrutyn E (eds) Bacterial infections of humans:
Ecological populations of bacteria act as socially epidemiology and control, 4th edn. Springer
cohesive units of antibiotic production and Science + Business Media, LLC, New York, NY,
resistance. Science 337:1228–1231 pp 543–575
4. Rossolini GM, Mantengoli E (2008) 7. Rigottier-Gois L, Alberti A et al (2011) Large-
Antimicrobial resistance in Europe and its scale screening of a targeted Enterococcus
f aecalis mutant library identifies envelope fitness 22. Prado-Prado FJ, Garcia-Mera X et al (2010)
factors. PLoS One 6:e29023 Multi-target spectral moment QSAR versus
8. Tenover FC, McGowan JE Jr (2009) The epi- ANN for antiparasitic drugs against different
demiology of bacterial resistance to antimicro- parasite species. Bioorg Med Chem 18:
bial agents. In: Brachman PS, Abrutyn E (eds) 2225–2231
Bacterial infections of humans: epidemiology 23. Garcia I, Fall Y et al (2011) First computational
and control, 4th edn. Springer Science + Business chemistry multi-target model for anti-
Media, LLC, New York, NY, pp 91–104 Alzheimer, anti-parasitic, anti-fungi, and anti-
9. Feuerriegel S, Oberhauser B et al (2012) bacterial activity of GSK-3 inhibitors in vitro,
Sequence analysis for detection of first-line in vivo, and in different cellular lines. Mol
drug resistance in Mycobacterium tuberculosis Divers 15:561–567
strains from a high-incidence setting. BMC 24. Speck-Planche A, Kleandrova VV et al (2012)
Microbiol 12:90 Fragment-based approach for the in silico dis-
10. Lienhardt C, Glaziou P et al (2012) Global covery of multi-target insecticides. Chemometr
tuberculosis control: lessons learnt and future Intell Lab Syst 111:39–45
prospects. Nat Rev Microbiol 10:407–416 25. Speck-Planche A, Kleandrova VV et al (2012)
11. Hann MM, Oprea TI (2004) Pursuing the In silico discovery and virtual screening of
leadlikeness concept in pharmaceutical multi-target inhibitors for proteins in
research. Curr Opin Chem Biol 8:255–263 Mycobacterium tuberculosis. Comb Chem
12. Lazo JS, Wipf P (2000) Combinatorial chemistry High Throughput Screen 15:666–673
and contemporary pharmacology. J Pharmacol 26. Speck-Planche A, Kleandrova VV et al (2012)
Exp Ther 293:705–709 Chemoinformatics in anti-cancer chemother-
13. Bleicher KH, Bohm HJ et al (2003) Hit and apy: multi-target QSAR model for the in silico
lead generation: beyond high-throughput discovery of anti-breast cancer agents. Eur J
screening. Nat Rev Drug Discov 2:369–378 Pharm Sci 47:273–279
14. Hansch C, Leo A (1995) Exploring QSAR: 27. Speck Planche A, Cordeiro MNDS (2013) In
fundamentals and applications in chemistry Chemoinformatics in drug design. Artificial
and biology. American Chemical Society, neural networks for simultaneous prediction
Washington, DC of anti-enterococci activities and toxicological
profiles. Proceedings of the 5th International
15. Jorgensen WL (2004) The many roles of
joint conference on computational intelli-
computation in drug discovery. Science 303: gence, NCTA-International conference on
1813–1818 neural computation theory and applications,
16. Oprea T (2005) Chemoinformatics in drug Vilamoura, Algarve, Portugal, 20–22 Sept,
discovery. WILEY-VCH Verlag GmbH & Co. pp 458–465
KGaA, Weinheim 28. Luan F, Cordeiro MNDS et al (2013) TOPS-
17. Brown N, Lewis RA (2006) Exploiting QSAR MODE model of multiplexing neuroprotective
methods in lead optimization. Curr Opin Drug effects of drugs and experimental-theoretic
Discov Devel 9:419–424 study of new 1,3-rasagiline derivatives poten-
18. Borchardt RT, Kerns EH et al (eds) (2006) tially useful in neurodegenerative diseases.
Optimizing the “drug-like” properties of leads Bioorg Med Chem 21:1870–1879
in drug discovery. Springer Science + Business 29. Tenorio-Borroto E, Penuelas Rivas CG et al
Media, LLC, New York, NY (2012) ANN multiplexing model of drugs
19. Croes S, Koop AH et al (2012) Efficacy, nephro- effect on macrophages; theoretical and flow
toxicity and ototoxicity of aminoglycosides, math- cytometry study on the cytotoxicity of the anti-
ematically modelled for modelling- supported microbial drug G1 in spleen. Bioorg Med Chem
therapeutic drug monitoring. Eur J Pharm Sci 20:6181–6194
45:90–100 30. Speck-Planche A, Kleandrova VV et al (2013)
20. Hau J, Schapiro SJ (2011) Handbook of labo- New insights toward the discovery of antibac-
ratory animal science: essential principles and terial agents: multi-tasking QSBER model for
practices. CRC Press, Taylor & Francis Group, the simultaneous prediction of anti-tuberculosis
LLC, Boca Raton, FL activity and toxicological profiles of drugs. Eur
21. Vina D, Uriarte E et al (2009) Alignment-free J Pharm Sci 48:812–818
prediction of a drug-target complex network 31. Speck-Planche A, Kleandrova VV et al (2013)
based on parameters of drug connectivity and Chemoinformatics for rational discovery of
protein sequence of receptors. Mol Pharm safe antibacterial drugs: simultaneous predic-
6:825–835 tions of biological activity against streptococci
and toxicological profiles in laboratory ani- 47. Bhonsle JB, Venugopal D et al (2007)
mals. Bioorg Med Chem 21:2727–2732 Application of 3D-QSAR for identification of
32. Speck-Planche A, Cordeiro MNDS (2013) descriptors defining bioactivity of antimicrobial
Simultaneous modeling of antimycobacterial peptides. J Med Chem 50:6545–6553
activities and ADMET profiles: a chemoinfor- 48. Bucinski A et al (2004) Artificial neural net-
matic approach to medicinal chemistry. Curr works for prediction of antibacterial activity in
Top Med Chem 13:1656–1665 series of imidazole derivatives. Comb Chem
33. van de Waterbeemd H (1995) Chemometrics High Throughput Screen 7:327–336
methods in molecular design. VCH Publishers, 49. Todeschini R, Consonni V (2009) Molecular
Weinheim descriptors for chemoinformatics. WILEY-
34. Gaulton A, Bellis LJ et al (2012) ChEMBL: a VCH Verlag GmbH & Co. KGaA, Weinheim
large-scale bioactivity database for drug discov- 50. Estrada E, Matamala AR (2007) Generalized
ery. Nucleic Acids Res 40:D1100–D1107 topological indices. Modeling gas-phase rate
35. Knox C, Law V et al (2011) DrugBank 3.0: a coefficients of atmospheric relevance. J Chem
comprehensive resource for ‘omics’ research on Inf Model 47:794–804
drugs. Nucleic Acids Res 39:D1035–D1041 51. Estrada E, Uriarte E et al (2000) A novel
36. Mok NY, Brenk R (2011) Mining the ChEMBL approach for the virtual screening and rational
database: an efficient chemoinformatics work- design of anticancer compounds. J Med Chem
flow for assembling an ion channel- focused 43:1975–1985
screening library. J Chem Inf Model 51: 52. Roy K, Ghosh G (2004) QSTR with extended
2449–2454 topochemical atom indices. 2. Fish toxicity of
37. Todeschini R, Consonni V (2000) Handbook substituted benzenes. J Chem Inf Comput Sci
of molecular descriptors. WILEY-VCH Verlag 44:559–567
GmbH, Weinheim 53. Roy K, Ghosh G (2008) QSTR with extended
38. Kubinyi H (1993) QSAR: Hansch analysis and topochemical atom indices. 10. Modeling of
related approaches. VCH Publishers, Weinheim toxicity of organic chemicals to humans using
different chemometric tools. Chem Biol Drug
39. Kubinyi H, Folkers G et al (eds) (2002) 3D Des 72:383–394
QSAR in drug design: recent advances. Kluwer
Academic Publishers, New York 54. Castillo-Garit JA, Vega MC et al (2011)
Ligand-based discovery of novel trypanosomi-
40. Klein CD, Hopfinger AJ (1998) Pharma cidal drug-like compounds: in silico identifica-
cological activity and membrane interactions of tion and experimental support. Eur J Med
antiarrhythmics: 4D-QSAR/QSPR analysis. Chem 46:3324–3330
Pharm Res 15:303–311
55. Casañola-Martin GM, Marrero-Ponce Y et al
41. Vedani A, Dobler M (2002) 5D-QSAR: the (2010) Bond-based 2D quadratic fingerprints
key for simulating induced fit? J Med Chem in QSAR studies: virtual and in vitro tyrosinase
45:2139–2149 inhibitory activity elucidation. Chem Biol
42. Vedani A, Dobler M et al (2005) Combining Drug Des 76:538–545
protein modeling and 6D-QSAR. Simulating 56. Barigye SJ, Marrero-Ponce Y et al (2013)
the binding of structurally diverse ligands to the Event-based criteria in GT-STAF information
estrogen receptor. J Med Chem 48:3700–3703 indices: theory, exploratory diversity analysis
43. Carloni P, Alber F (eds) (2003) Quantum and QSPR applications. SAR QSAR Environ
medicinal chemistry. WILEY-VCH Verlag Res 24:3–34
GmbH & Co. KGaA, Weinheim 57. Barigye SJ, Marrero-Ponce Y et al (2013)
44. Li P, Yin J et al (2013) Synthesis, antibacterial Relations frequency hypermatrices in mutual,
activities, and 3D-QSAR of sulfone derivatives conditional and joint entropy-based informa-
containing 1,3,4-oxadiazole moiety. Chem tion indices. J Comput Chem 34:259–274
Biol Drug Des 82:546–556 58. Vazquez-Prieto S, Gonzalez-Diaz H et al
45. Lu X, Lv M et al (2012) Pharmacophore and (2013) A QSPR-like model for multilocus gen-
molecular docking guided 3D-QSAR study of otype networks of Fasciola hepatica in
bacterial enoyl-ACP reductase (FabI) inhibi- Northwest Spain. J Theor Biol 343C:16–24
tors. Int J Mol Sci 13:6620–6638 59. Alonso N, Caamano O et al (2013) Model for
46. Uddin R, Lodhi MU et al (2012) Combined high-throughput screening of multi-target
pharmacophore and 3D-QSAR study on a drugs in chemical neurosciences; synthesis,
series of Staphylococcus aureus Sortase A assay and theoretic study of rasagiline carba-
inhibitors. Chem Biol Drug Des 80:300–314 mates. ACS Chem Neurosci 4:1393–1403
60. Estrada E, Molina E et al (2001) Can 3D descriptors and fingerprints. J Comput Chem
structural parameters be predicted from 2D 32:1466–1474
(topological) molecular descriptors? J Chem 72. Todeschini R, Lasagni M et al (1994) New
Inf Comput Sci 41:1015–1021 molecular descriptors for 2D and 3D struc-
61. Estrada E (2002) Physicochemical interpreta- tures. Theory. J Chemometr 8:263–272
tion of molecular connectivity indices. J Phys 73. Hill T, Lewicki P (2006) STATISTICS meth-
Chem A 106:9085–9091 ods and applications. A comprehensive refer-
62. Molina E, Diaz HG et al (2004) Designing ence for science, industry and data mining.
antibacterial compounds through a topological StatSoft, Tulsa
substructural approach. J Chem Inf Comput 74. Suzuki K (ed) (2011) Artificial neural net-
Sci 44:515–521 works: methodological advances and biomedi-
63. Gonzalez-Diaz H, Torres-Gomez LA et al cal applications. InTech, Rijeka
(2005) Markovian chemicals “in silico” design 75. Sabet R, Fassihi A et al (2012) Computer-aided
(MARCH-INSIDE), a promising approach for design of novel antibacterial 3-hydroxypyridine-
computer-aided molecular design III: 2.5D 4-ones: application of QSAR methods based on
indices for the discovery of antibacterials. J Mol the MOLMAP approach. J Comput Aided Mol
Model 11:116–123 Des 26:349–361
64. Marrero-Ponce Y, Marrero RM et al (2006) 76. Garcia-Domenech R, de Julian-Ortiz JV
Non-stochastic and stochastic linear indices of (1998) Antimicrobial activity characterization
the molecular pseudograph’s atom-adjacency in a heterogeneous group of compounds.
matrix: a novel approach for computational in J Chem Inf Comput Sci 38:445–449
silico screening and “rational” selection of 77. Lata S, Sharma BK et al (2007) Analysis and
new lead antibacterial agents. J Mol Model prediction of antibacterial peptides. BMC
12:255–271 Bioinformatics 8:263
65. Marrero-Ponce Y, Medina-Marrero R et al 78. Hall M, Frank E et al (2009) The WEKA data
(2005) Atom, atom-type, and total nonsto- mining software: an update. SIGKDD Explor
chastic and stochastic quadratic fingerprints: a 11:10–18
promising approach for modeling of antibacte- 79. Hall M, Frank E et al (1999–2013)
rial activity. Bioorg Med Chem 13:2881–2899 WEKA. Waikato Environment for Knowledge
66. Speck-Planche A, Scotti MT et al (2009) Analysis. v3.6.9, Hamilton
Design of novel antituberculosis compounds 80. Witten IH, Frank E et al (2011) Data mining:
using graph-theoretical and substructural practical machine learning tools and tech-
approaches. Mol Divers 13:445–458 niques. Morgan Kaufmann Publishers, Elsevier,
67. Estrada E (1996) Spectral moments of the Amsterdam
edge adjacency matrix in molecular graphs. 1. 81. StatSoft (2001) STATISTICA 6.0. Data analy-
Definition and applications for the prediction sis software system
of physical properties of alkanes. J Chem Inf 82. González-Díaz H, Pérez-Bello A et al (2007)
Comput Sci 36:844–849 Chemometrics for QSAR with low sequence
68. Estrada E (1997) Spectral moments of the homology: Mycobacterial promoter sequences
edge adjacency matrix in molecular graphs. recognition with 2D-RNA entropies.
2. Molecules containing heteroatoms and Chemometr Intell Lab Syst 85:20–26
QSAR applications. J Chem Inf Comput Sci 83. Hanczar B, Hua J et al (2010) Small-sample pre-
37:320–328 cision of ROC-related estimates. Bioinformatics
69. Estrada E (1998) Spectral moments of the 26:822–830
edge adjacency matrix in molecular graphs. 3. 84. Sader HS, Biedenbach DJ et al (2012)
Molecules containing cycles. J Chem Inf Antimicrobial activity of the investigational pleu-
Comput Sci 38:23–27 romutilin compound BC-3781 tested against
70. Estrada E, Gutiérrez Y (2002–2004) Gram-positive organisms commonly associated
MODESLAB. v1.5, Santiago de Compostela with acute bacterial skin and skin structure
71. Yap CW (2011) PaDEL-descriptor: an open infections. Antimicrob Agents Chemother 56:
source software to calculate molecular 1619–1623
Chapter 5
Use of Artificial Neural Networks in the QSAR Prediction

of Physicochemical Properties and Toxicities for REACH
Legislation
John C. Dearden and Philip H. Rowe
Abstract
With the introduction of the REACH legislation in the European Union, there is a requirement for property
and toxicity data on chemicals produced in or imported into the EU at levels of 1 tonne/year or more.
This has meant an increase in the in silico prediction of such data. One of the chief predictive approaches
is QSAR (quantitative structure–activity relationships), which is widely used in many fields.
A QSAR approach that is increasingly being used is that of artificial neural networks (ANNs), and this
chapter discusses its application to the range of physicochemical properties and toxicities required by
REACH. ANNs generally outperform the main QSAR approach of multiple linear regression (MLR),
although other approaches such as support vector machines sometimes outperform ANNs. Most ANN
QSARs reported to date comply with only two of the five OECD Guidelines for the Validation of (Q)SARs.
Key words QSAR, Artificial neural networks, Physicochemical properties, Toxicity, REACH
1 Introduction
The European Union’s REACH legislation (concerned with the

Registration, Evaluation, Authorisation and restriction of
CHemicals) [1] became operational in 2008. From that date,
chemicals produced in or imported into the EU at more than 1
tonne/year have to be registered based on the characterization of
their toxicities and physicochemical properties. However, com-
plete evaluation of these properties by experimental tests can be
time-consuming and expensive. Hence the REACH legislation
permits the use of (Q)SARs ((quantitative) structure–activity rela-
tionships) for toxicity and property prediction where appropriate.
A QSAR is a mathematical equation that correlates a toxicity or
property for a series of compounds with one or more structural prop-
erties of the compounds; that equation can then be used to predict
the toxicity or property of another (usually related) compound.
65
66 John C. Dearden and Philip H. Rowe
It may seem strange that a complex endpoint such as a toxicity can be

modeled in such a way, but it should be remembered that all molecu-
lar properties—chemical, physical, and biological—are a function of
molecular structure. An example of a simple QSAR is Eq. 1, which
models the toxicity of barbiturates to the mouse [2]:
Log (1 / LD50 ) = 1.02 log P − 0.27 ( log P ) + 1.86

2
(1)

= =
n 13 R 2 0.852 s = 0.113
In this QSAR, LD50 is the dose required to kill 50 % of the mice,
and P is the octanol–water partition coefficient, an important
property that reflects the ability of compounds to penetrate lipid
membranes in an organism; n is the number of compounds used to
develop the QSAR; R2 is the coefficient of determination, the frac-
tion of the toxicity that is described by the equation; and s is the
standard error of the prediction. It can be seen that the partition
coefficient, a simple physicochemical property, models 85.2 % of
the acute toxicity of barbiturates to the mouse.
Another example, shown in Eq. 2, concerns the toxicity of
nitrobenzenes to the aquatic ciliate Tetrahymena pyriformis [3]:
Log (1 / IGC50 ) = 0.467 log P − 1.60 E LUMO − 2.55 (2)

= =
n 42 R 2 0.881 s = 0.246
IGC50 is the dose that reduces growth rate of the ciliate by 50 %.
ELUMO is the energy of the lowest unoccupied molecular orbital, a
measure of the electron-attracting ability (electrophilicity) of a
compound. Equation 2 demonstrates the use of both a hydropho-
bic descriptor property, P, and an electronic property, ELUMO, in
the same equation.
Equations 1 and 2 are examples of multiple linear regression
(MLR), which is the most widely used QSAR modeling technique.
There are, however, numerous other techniques that are used in
QSAR modeling, including artificial neural networks (ANNs) [4],
nonlinear regression, partial least squares, genetic algorithms, and
support vector machines (SVM) [5, 6]. Each has its own strengths
and weaknesses.
The main strength of ANNs is that they are nonlinear (strictly,
nonrectilinear) methods and so can deal with structure–activity
relationships in which there are nonrectilinear correlations between
activity and one or more descriptor properties. In Eq. 1 above, there
is a nonrectilinear correlation between barbiturate toxicity and parti-
tion coefficient. In this case, the correlation is parabolic, which can
be accounted for by the incorporation of a (log P)2 term. However,
in many cases, the form of nonrectilinearity is not known, so an
MLR approach will not yield good correlations. This is particu-
larly important when non-congeneric data sets are being modeled.
Artificial Neural Networks for REACH Property/Toxicity 67
It should be noted that most ANN-based QSAR studies have used a

back-propagation algorithm [7].
The main weaknesses of ANNs are that they do not yield a
QSAR equation directly, they are difficult to interpret, and they do
not work well with small data sets [7].
It is important to know the REACH requirements and restric-
tions concerning the use of QSARs, as stated by its Annex XI:
Results obtained from valid qualitative or quantitative structure-activity
relationship models ((Q)SARs) may indicate the presence or absence
of a certain dangerous property. Results of (Q)SARs may be used
instead of testing when the following conditions are met:
–– results are derived from a (Q)SAR model whose scientific validity
has been established,
–– the substance falls within the applicability domain of the (Q)SAR
model,
–– results are adequate for the purpose of classification and labelling
and/or risk assessment, and,
–– adequate and reliable documentation of the applied method is
provided.
In addition, guidance on the use of (Q)SARs for the imple-
mentation of REACH has been provided by the European
Chemicals Agency (ECHA) [8], which states that:
In principle, (Q)SARs can be applied in a number of ways, namely to:
(a) provide information for use in priority setting procedures;
(b) guide the experimental design of an experimental test or testing
strategy;
(c) improve the evaluation of existing test data;
(d) provide mechanistic information (which could be used, for example,
to support the grouping of chemicals into categories);
(e) fill a data gap needed for hazard and risk assessment;
(f) fill a data gap needed for classification and labelling;
(g) fill a data gap needed for PBT or vPvB assessment.
The first four applications (a–d) are more general regulatory applica-
tions of QSARs, whereas the last three applications (e–g) are more
REACH-specific.
In some situations, (Q)SARs could be used to replace test data,
whereas in other situations, the models would be used to provide sup-
plementary information to experimental data. In practice, it is foreseen
that (Q)SAR information will most often be used to supplement
experimental test data within chemical categories and endpoint-specific
Integrated Testing Strategies (ITS). However, it is expected that (Q)
SARs will be used increasingly for the direct replacement of test data,
as relevant and reliable models become increasingly available, and as
experience in their use becomes more widespread.
The acronym QSAR (quantitative structure–activity relation-
ship) in REACH documentation is taken to include QSPR (quan-
titative structure–property relationship).
To be used in a regulatory context, QSAR models should have,

according to the five principles proposed by the Organisation for
Economic Co-operation and Development (OECD) [9, 10]:
(a) A defined endpoint
(b) An unambiguous algorithm
(c) A defined domain of applicability
(d) Appropriate measures of goodness of fit, robustness, and
predictivity
(e) A mechanistic interpretation, if possible
In particular, a series of internal and external validation tests
are used to demonstrate the reliability of a QSAR model. Goodness
of fit is evaluated for the compounds used to derive the model
(i.e., the compounds of the training set). Robustness is estimated
by cross validation and/or randomization techniques. Predictive
power is generally evaluated using an external validation set of
compounds that were not used to develop the model, and it is
now a standard requirement that external validation of QSAR
models is performed.
A useful review of QSAR modeling for regulatory purposes is
given by Fjodorova et al. [11].
2 The Value of ANN-Derived QSARs
In general, the neural network approach to QSAR development has

several important advantages over the standard multiple linear
regression (MLR) approach [12]. Firstly, as pointed out above, the
ability of ANNs to utilize nonrectilinear functions can result in bet-
ter models being obtained. Secondly, some ANNs are able to deal
with interactions between descriptors. Thirdly, they are able to han-
dle very noisy data, and fourthly, as a corollary to the above, they
are valuable for the derivation of QSARs for non-congeneric series.
There is a plethora of ANN methods available for QSAR mod-
eling [13–15], such as back-propagation NNs, counterpropaga-
tion NNs, probabilistic NNs, radial basis function NNs, and
computational NNs [16]; Baskin et al. [13] list 35 types of ANNs
that have been used in QSAR studies. Care must be taken to select
the most appropriate method and to guard against overtraining of
the network, which lowers the predictivity of the derived model.
It should be noted that in comparison with other methods for
the derivation of QSAR models, ANN methods are not always the
best performers. A study of the prediction of biomagnification factors
for organochlorine compounds [17] found that a feed-forward ANN
approach gave better results than did an MLR approach. However,
for the prediction of acute toxicity of diverse organic chemicals to the
fathead minnow [18], a support vector machine (SVM) approach
gave better results than did a back-propagation ANN.
3 Do ANN-Derived QSARs Meet Validity Requirements for Regulatory Purposes?
Clearly, ANN-derived QSARs must have a defined endpoint in

order to satisfy the first OECD principle, and there is no greater
difficulty with that than there is with any other QSAR approach. It
must be stressed, however, that rigor is required in assembling and
checking the database used to develop any QSAR [19].
The second OECD principle requires an unambiguous algo-
rithm, which cannot be obtained with ANN techniques, since they
are generally regarded as “black-box” methods. The descriptors
used to build an ANN model are known, since they form most of
the input neuron layer, but their signs and coefficients are not
known. Hence the algorithm cannot be said to be unambiguous.
However, Baskin et al. [20] and Jurs and coworkers [21, 22] have
proposed ways in which the relative importance of the descriptors
in an ANN-derived QSAR can be obtained. This means [13] that
“the interpretability of neural network models is not reduced in
comparison with traditional statistical linear approaches, at least for
interpretation methods based on ranking the relative importance of
descriptors.” However, these proposals appear to have been little
used to date.
The third OECD principle requires a defined applicability
domain, which to a first approximation represents the chemical
and biological space covered by the chemicals in the training set,
irrespective of the QSAR approach used [23].
The fourth OECD principle calls for appropriate measures of
goodness of fit, robustness, and predictivity, and these are as read-
ily available for ANN models as for other types of QSAR models.
It should be mentioned, however, that great care must be taken (a)
to avoid overtraining in the development of an ANN QSAR and
(b) to ensure its external validation.
The fifth OECD principle calls for a mechanistic interpretation
of the QSAR model, if possible. Thus there is a recognition that
such interpretations cannot always be achieved. Johnson [24] has
pointed out that “QSAR has devolved into a perfectly practiced art
of logical fallacy; cum hoc ergo propter hoc (with this, therefore
because of this).” He also commented that a statement to the
effect that, say, a particular descriptor represents molecular shape is
not a mechanistic interpretation. It may also be noted that there is
currently a commendable increase in the number of QSARs pub-
lished that are derived from compounds known to act by a given
mechanism. In such cases there is no (or less) need to interpret the
QSAR mechanistically. An example is the identification of different
mechanisms of action of chemicals that are skin sensitizers, prior to
QSAR analysis [25].
Following publication of the OECD principles [8], Vračko
et al. [26] carried out a validation study of counterpropagation
ANN modeling of fish toxicity. While acknowledging the potential
drawbacks of the method, they concluded that “a (Q)SAR model

can be derived and validated using a counter propagation NN
approach whilst still satisfying most if not all of the OECD princi-
ples for validation of (Q)SAR models.”
However, this does not necessarily mean that ANN-derived
QSARs are automatically acceptable for regulatory purposes,
including those relating to the REACH legislation. A number of
peer-reviewed QSARs are stored in the (Q)SAR Model Reporting
Format (QMRF) database at the European Commission Joint
Research Centre (JRC) at Ispra, Italy [27]. The website [28] states:
“The database is intended to help to identify valid (Q)SARs, e.g.
for the purposes of REACH…The information is structured
according to the OECD principles for the validation of (Q)SAR
models.” Nevertheless, it is important to note that the quality and
relevance of QSARs in the QMRF database are not checked by the
JRC before inclusion.
There are, at the time of writing, 70 QSARs in the QMRF
database, of which 13 are ANN-derived QSARs, covering a range
of toxicity endpoints such as acute toxicity to Daphnia magna,
fathead minnow, birds, and rat (oral), mouse carcinogenicity, der-
mal irritation, skin sensitization, and chromosomal aberration, as
well as QSARs for biodegradation and octanol–water partition
coefficient.
It should be noted that QSARs for the prediction of physico-
chemical properties are usually referred to as QSPRs (quantitative
structure–property relationships).
4 Some Case Studies of ANN-Derived QSARs
As pointed out above, QSARs to be used in connection with

REACH need to comply with most of the OECD principles for
validation of (Q)SAR models. Hence in this chapter an indication
is given of that compliance; for example, (OECD 1, 2, 4I,E)
indicates that a model complies with OECD principles 1, 2, and 4,
with I and E indicating that there is both internal and external vali-
dation of the model.
4.1 Physicochemical The REACH legislation [29] requires information on up to 19

Properties physicochemical properties, depending on annual supply levels.
More detailed information can be found in Regulation (EC) No.
1907/2006 [30].
From Annex VII (required for substances at a supply level of
≥1 tonne/year):
State of the substance, melting/freezing point, boiling point,
relative density, vapor pressure, surface tension, water solubility,
n-octanol–water partition coefficient, flash point, flammability*,
explosive properties, self-ignition temperature, oxidizing proper-
ties, and granulometry (particle size distribution).
*Flammability includes “pyrophoric properties,” “flammability

on contact with water,” and “flammability upon ignition.”
From Annex VIII (additionally required for substances at a
supply level of ≥10 tonnes/year):
Adsorption/desorption (generally taken to mean relating to
soil).
From Annex IX additionally (required for substances at a sup-
ply level of ≥100 tonnes/year):
Dissociation constant, viscosity, and stability in organic solvents.
In addition, the ECHA endpoint-specific guidance [29]
includes an appendix on the determination of the air–water parti-
tion coefficient, otherwise known as Henry’s law constant. This is
an important property with applications in, for example, inhalation
toxicology and environmental distribution of chemicals.
Of the above properties, two (state of the substance and gran-
ulometry) are not amenable to QSPR analysis, and for two others
(oxidizing properties and stability in organic solvents), QSPRs
have not yet been derived.
Dearden et al. [31] have discussed the in silico prediction of all
the above properties that are amenable to QSPR analysis, and
Taskinen and Yliruusi [32] and Devillers [14] have reviewed the
ANN QSPR modeling of a number of those properties.
4.1.1 Melting/ Godavarthy et al. [33] used a back-propagation NN to model

Freezing Point the melting points of over 1,200 organic chemicals, with
R2 = 0.99 and RMSE (root mean square error) = 12.6° (OECD
compliance 1,4IE) for the training set and 10.8° for the test set.
A typical melting point prediction error is around 30°–40°, so
this result is surprisingly good. However, the statistics are per-
haps too good, suggesting that over-fitting may have taken place.
More realistic results were obtained [34] for a data set of 4,173
organic chemicals using a feed-forward back-propagation NN
and 2- and 3-dimensional (2D and 3D) descriptors; the use of
2D descriptors gave the better results, with a training-set mean
absolute error (MAE) of 37.6° and a 277-compound test-set
MAE of 32.6° (OECD compliance 1,3,4IE,5).
It should be noted that melting point is probably the least well
predicted physicochemical property, partly because of the difficulty
of modeling crystal packing.
4.1.2 Boiling Point Boiling point is somewhat easier to predict. Hall and Story [35]
used a back-propagation NN with atom-type electrotopological
descriptors to model the boiling points of 298 diverse organic
chemicals, with an MAE of 3.9° (OECD compliance 1,4IE). Using
a feed-forward NN with quantum-mechanical descriptors, Chalk
et al. [36] modeled the boiling points of a large number of organic
chemicals, with a 6,000-compound training-set standard deviation
of 16.5° and a 629-compound test-set standard deviation of 19.0°
(OECD compliance 1,4IE). They commented that the low error

found by Hall and Story [35] could have been caused by overtrain-
ing of their model.
4.1.3 Relative Density Relative density is quite simple to calculate, for liquids at least [31].
Gakh et al. [37] used a feed-forward back-propagation NN to pre-
dict the densities of 134 liquid hydrocarbons at 25 °C, using
descriptors derived from graph theory. They found an MAE of
0.6 % for a 25-chemical external test set (OECD compliance 1,4E).
Group contributions and a feed-forward back-propagation NN
with two hidden layers were used to predict the densities of 82
ionic liquids [38]. An external test set of 24 ionic liquids yielded an
MAE of 0.26 % (OECD compliance 1,4E).
4.1.4 Vapor Pressure Vapor pressure, being temperature dependent, must be reported
for a specific temperature, such as 25 °C (298 K). McClelland and
Jurs [39] used a computational NN to model the vapor pressures
at 25 °C of 420 diverse organic chemicals. They obtained RMSE
values of 0.19 and 0.33 log unit for the training and validation sets,
respectively (OECD compliance 1,4IE). Chalk et al. [40] used a
feed-forward NN to model the variation of vapor pressure with
temperature, using quantum-mechanical descriptors with a data
set of 7,681 measurements for 2,349 chemicals. Their standard
deviations were 0.322 and 0.326 log unit for training and valida-
tion sets, respectively (OECD compliance 1,4IE).
4.1.5 Surface Tension The surface tensions at various temperatures of 82 organic liquids
were modeled with a multilayer perceptron NN, with five physical
properties as descriptors [41], and with an MAE of 1.41 % for the
training set of 734 data points and 1.95 % for the test set of 314
data points (OECD compliance 1,4E). Gharagheizi et al. [42] used
a much larger number (752) of organic liquids and a range of tem-
peratures. Using a feed-forward NN with group contribution
descriptors, they obtained standard deviations of 1.7 % for training
and internal and external validation sets (OECD compliance 1,4IE).
4.1.6 Water Solubility One of the most important physicochemical properties is aqueous
solubility, and there have been many QSPR studies carried out in
this area [43], including those using ANN approaches. Yan and
Gasteiger [44] used a back-propagation NN with 3D descriptors
to model a set of 1,293 organic chemicals. They obtained, for the
797-compound training set and the 496-compound test set, MAEs
0.41 and 0.49 log unit (OECD compliance 1,4IE). These com-
pare well with the typical experimental error in aqueous solubility
of about 0.58 log unit [45]. They are also considerably better than
an MLR model for the same chemicals (MAEs 0.70 and 0.68 log
unit for training and test sets, respectively). However, when an
external set of 1,587 Merck chemicals was tested, the NN MAE
was 0.77 log unit. The authors commented that this was probably
due to greater chemical diversity of the Merck chemicals. Similar
results were obtained by Bruneau [46] using a Bayesian NN with
physicochemical descriptors for a training set of 1,560 organic
chemicals (standard error = 0.53 log unit) and an external test set
of 934 organic chemicals (standard error = 0.81 log unit) (OECD
compliance 1,4IE).
4.1.7 n-Octanol–Water Without doubt the most important physicochemical property for
Partition Coefficient modeling bioactivity is the n-octanol–water partition coefficient
(P, Kow). In QSAR studies it is always used in logarithmic form as
log P or log Kow. Some 70 % of all QSARs include this term or one
related to it, such as the chromatographic retention term Rm. Its
importance lies in the fact that it is a surrogate for lipid–water par-
titioning and hence models xenobiotic transport in an organism,
although it can contribute also to receptor binding. There are a
number of reviews of the topic [47, 48] and many published stud-
ies. Chen [49] compared the performance of MLR, radial basis
function neural networks, and support vector machines for the
prediction of log P of about 3,500 organic chemicals. The SVM
results were best (training set s = 0.54, test set s = 0.56), with MLR
slightly worse (training set s = 0.62, test set s = 0.56) and RBFNN
somewhat worse (training set s = 0.56, test set s = 0.72) (OECD
compliance 1,4IE). Tetko et al. [50] used electrotopological state
(E-state) indices [51] with an ensemble of 50 different neural net-
works to model log P values of a very large number of organic
chemicals; for a 12,908-compound training set they obtained
RMSE =
0.38 log unit, and for a 1,174-compound test set
RMSE = 0.36 log unit (OECD compliance 1,4IE). These results
were better than those from MLR and from several commercially
available software programs. Considering the very large number of
chemicals used, this is a very good result.
4.1.8 Flash Point The flash point of a chemical is the temperature at which the vapor
concentration is equal to the lower explosive limit. Katritzky et al.
[52] used both MLR and back-propagation NN with CODESSA
descriptors to model the flash points of up to 758 organic chemi-
cals. For the training sets, MLR yielded MAE = 13.9°, while NN
yielded better results: MAE = 12.6° (OECD compliance 1,4IE).
No test-set results were reported. Gharagheizi et al. [53], using a
feed-forward NN and group contributions as descriptors with a
large data set of organic chemicals, found MAE = 7.9° for the
1,241-compound training set and MAE = 9.9° for the
137-compound test set (OECD compliance 1,4IE).
4.1.9 Flammability Upper and lower flammability limits represent the range of concen-
Limits trations in air within which a chemical will ignite provided that an
ignition source is present. Studies of both limits have been made by
Gharagheizi. He used a feed-forward NN with group contributions

to model the upper flammability limits of 867 organic chemicals
[54]; for the 694-compound training set MAE = 7.26 % and for the
173-compound test set MAE = 6.23 % (OECD compliance 1,4IE),
whereas using MLR on the same data set [55], he obtained
MAE = 9.2 % for the full data set (MAEs not given for training and
test sets) (OECD compliance 1,2,4IE).
In similar fashion Gharagheizi investigated the lower flamma-
bility limits of a data set of 1,057 organic chemicals. Using MLR
[56] he obtained MAE = 7.68 % for the full data set (MAEs not
given for training and test sets) (OECD compliance 1,2,4IE),
while with a feed-forward NN [57], Gharagheizi obtained
MAE = 4.35 % for the 846-compound training set and MAE = 5.70 %
for the 211-compound test set (OECD compliance 1,4IE). Clearly,
for the prediction of both upper and lower flammability limits, the
neural network approach yielded considerably better results.
4.1.10 Explosive Explosive properties have to date been investigated very little by
Properties QSAR, with the exception of impact sensitivity. Cho et al. [58] used
back-propagation NN with structural, topological, and physico-
chemical descriptors to model the impact sensitivity of 263 nitro
compounds, with a standard error of prediction of 0.211 log unit
(standard errors not given for training and test sets) (OECD com-
pliance 1,4IE). Wang et al. [59] compared the performance of
MLR, partial least squares (PLS), and back-propagation NN in
modeling the impact sensitivity of 156 diverse nitro compounds.
The RMSEs found for the 127-compound training set and the
49-compound test set, respectively, were MLR 0.210, 0.251; PLS
0.214, 0.250; and BPNN 0.192, 0.247 (OECD compliance 1,4IE).
Clearly the back-propagation NN approach gave the best results.
4.1.11 Self-Ignition MLR, PLS, and radial basis function and back-propagation NNs
Temperature (Autoignition with structural and physicochemical descriptors were used by
Temperature) Tetteh et al. [60] to model the self-ignition temperature of 233
organic chemicals. Overall MAEs were MLR 90.2°, PLS 38.4°,
RBFNN 30.1°, and BPNN 29.9° (MAEs not given for training
and test sets separately) (OECD compliance 1,4IE). Pan et al. [61]
used MLR and back-propagation NNs with E-state descriptors and
group contributions, respectively, to model the self-ignition tem-
peratures of 118 hydrocarbons. MAEs for the 42-compound test
set were MLR 32.4°, E-state BPNN 21.6°, and group contribu-
tion BPNN 24.8° (OECD compliance 1,4E).
4.1.12 Adsorption/ For REACH purposes, adsorption/desorption relates mainly to

Desorption sorption on soil and also on sediment and sludge. Liu and Yu [62]
modeled the sorption of 42 anilines and phenols on soil, using
both MLR and back-propagation NN. With log P, quantum-
chemical, and topological descriptors, they found standard errors
of prediction of 0.374 log unit for MLR and 0.282 log unit for
BPNN (OECD compliance 1,4I). No external validation was per-
formed. Soil sorption of 124 pesticides was investigated by
Goudarzi et al. [63] using SPA-MLR and SPA-ANN with physico-
chemical and structural descriptors (SPA is successive projection
algorithm); the type of NN used was not specified. The RMSEs
obtained (log units) were training set 0.420 (MLR) and 0.282
(NN) and test set 0.371 (MLR) and 0.289 (NN) (OECD compli-
ance 1,4IE).
4.1.13 Dissociation Dissociation constant is a key property of many chemicals, and

Constant hence much effort has gone into developing methods of predicting
pKa values. Luan et al. [64] investigated both MLR and radial
basis function NN approaches to pKa prediction of 74 drugs, using
CODESSA descriptors [65]. The neural network approach gave
the better results, with MLR yielding RMSE = 0.482 (training set)
and 0.987 (test set) and RBFNN yielding RMSE = 0.458 (training
set) and 0.613 (test set) (OECD compliance 1,4IE). Habibi-
Yangjeh et al. [66], using MLR and feed-forward back-propagation
NN with physicochemical descriptors on 242 benzoic acids and
phenols, also found much better performance with the neural net-
work approach. RMSE values (log units) were MLR 0.86
(205-compound training set) and 0.94 (37-compound test set)
and BPNN 0.26 (training set) and 0.27 (test set) (OECD compli-
ance 1,4IE).
4.1.14 Viscosity Using their ADAPT descriptors, Kauffman and Jurs [67] devel-
oped both MLR and computational NN models to predict the
viscosity of almost 200 organic solvents. For 170 training-set
chemicals, RMSE (MLR) was found to be 0.257 mPa s, while
RMSE (CNN) was 0.147 mPa s. RMSEs for the 21-compound
test set were (MLR) 0.278 and (CNN) 0.242 mPa s (OECD com-
pliance 1,4IE). Artemenko et al. [68] found a back-propagation
NN, with molecular fragment descriptors, to yield better predic-
tions than did MLR for a diverse set of liquid organic chemicals.
For 293 training-set chemicals, RMSEs (log units) were MLR
0.111 and BPNN 0.212. RMSEs for the 37-compound test set
were MLR 0.212 and BPNN 0.141 (OECD compliance 1,4IE).
4.1.15 Air–Water Although information on Henry’s law constant (HLC) is not a

Partition Coefficient REACH requirement, its documentation [29] discusses the
(Henry’s Law Constant) measurement and prediction of HLC, so it is pertinent to con-
sider ANN approaches to its prediction. Modarresi et al. [69]
compared the performance of MLR and radial basis function NN
for prediction of HLC values of 940 diverse organic chemicals
using a wide range of physicochemical and structural descriptors.
The neural network approach gave slightly better RMSE values
(log units) (training set = 0.564, test set 0.520) than did MLR
(training set = 0.570, test set 0.520) (OECD compliance 1,4IE,5).

Gharagheizi et al. [70] used a group contribution method with a
feed-forward NN to predict HLC values of a large group of organic
chemicals. RMSE values (log units) were 1,649-compound train-
ing set = 0.08 and 97-compound test set = 0.12 (OECD compli-
ance 1,3,4E).
4.2 Toxicological The REACH legislation [71] requires information on a total of 19

Properties toxicological properties, depending on annual supply levels. Some
of the properties relate to human health, and some to ecotoxicity.
More detailed information can be found in Regulation (EC) No.
1907/2006 [30].
From Annex VII (required for substances at a supply level of
≥1 tonne/year):
Skin irritation or corrosion, eye irritation, skin sensitization,
mutagenicity, and acute toxicity (oral if inhalation data are not
available).
Aquatic toxicity (Daphnia preferred), growth inhibition of
aquatic plants, and degradation (ready biodegradability).
From Annex VIII (additionally required for substances at a
Acute toxicity (dermal, if inhalation is unlikely), short-term
repeated-dose toxicity, reproductive toxicity, and toxicokinetics if
available.
Short-term fish toxicity, activated sludge respiration inhibition,
abiotic degradation (hydrolysis), and adsorption/desorption (gen-
erally taken to mean relating to soil).
Note: adsorption/desorption is also a requirement under phys-
icochemical properties and is covered in Subheading 4.1.12 above.
From Annex IX (additionally required for substances at a
Sub-chronic rodent toxicity, prenatal developmental toxicity,
and two-generation reproductive toxicity.
Long-term aquatic toxicity, bioaccumulation in aquatic spe-
cies, and effects on terrestrial organisms (invertebrates, soil micro-
organisms, plants).
From Annex X (additionally required for substances at a supply
level of ≥1,000 tonnes/year):
Carcinogenicity, long-term toxicity to invertebrates, sediment
microorganisms, plants, and birds.
Note: Annex X also states that additional information may be
required on some toxicity endpoints listed under Annexes VII–IX.
4.2.1 Skin Irritation Salina et al. [72] have reviewed the QSAR prediction of skin irritation
or Corrosion and corrosion. Golla et al. [73] used a feed-forward back-propagation
NN with physicochemical and structural descriptors to model the
rabbit skin irritation of 186 organic chemicals and obtained
RMSE = 1.1 unit, which is reasonable for a data range of 0–8 units.
An external test set of 22 chemicals was also tested, but no RMSE

value was given (OECD compliance 1,4IE). Skin corrosivity was
studied by Barratt [74] using a back-propagation NN with a small
number of physicochemical properties for classification as corro-
sive or noncorrosive; for 50 organic acids, 40 organic bases, and
33 phenols, he found 97.02, 93.86, and 95.85 % correct classifica-
tion (OECD compliance 1,4). No external test set was used.
4.2.2 Eye Irritation Salina et al. [72] have reviewed the QSAR prediction of eye irrita-
tion. Barratt [75] investigated the eye irritation of 57 aliphatic
alcohols and other neutral organic chemicals using a back-
propagation NN with a small number of physicochemical proper-
ties for classification as irritant or nonirritant and obtained 97.37 %
correct classification (OECD compliance 1,4). No external test set
was used.
Patlewicz et al. [76] investigated a set of 19 cationic surfactants
with back-propagation NN using a small number of physicochemi-
cal properties, which yielded R2 = 0.702 (OECD compliance 1,4).
No error values were given.
4.2.3 Skin Sensitization Patlewicz et al. [77] have reviewed the QSAR prediction of skin
sensitization. Devillers [78] used a feed-forward back-propagation
NN with structural and physicochemical descriptors in a classifica-
tion model of the skin sensitization of 259 organic chemicals and
obtained an overall correct classification of 89.19 %, compared
with 82.63 % using linear discriminant analysis [79] (OECD com-
pliance 1,4IE). No external test-set results were given.
Golla et al. [80] also used a feed-forward back-propagation
NN with structural and physicochemical descriptors in a classifica-
tion model of skin sensitization, using three different endpoints
(mouse local lymph node assay (LLNA), guinea-pig maximization
test (GPMT), and a test from the German Federal Institute for
Health Protection of Consumers and Veterinary Medicine
(BgVV)). The training-set results were LLNA (358 chemicals)
90 % correct classification, GPMT (307 chemicals) 95 %, and
BgVV (251 chemicals) 90 % (OECD compliance 1,4IE). No test-
set results were given.
4.2.4 Mutagenicity Benigni [81] has reviewed the QSAR prediction of mutagenicity.
Xu et al. [82] investigated a very large data set of 7,617 organic
chemicals, of which 4,252 were mutagens and 3,365 were non-
mutagens. Using molecular fingerprints as descriptors, they
employed five different approaches, namely, SVM, decision tree,
feed-forward NN, k-nearest neighbor, and naïve Bayes, to classify
the chemicals. The prediction accuracies for the training set were,
respectively, 84.1 %, 80.4 %, 80.2 %, 82.1 %, and 65.8 % (OECD
compliance 1,4). The NN method, while good, was not quite as
accurate as the SVM, decision tree, and k-nearest neighbor meth-
ods. No test-set results were given.
For a set of aromatic amines, Leong et al. [83] also employed

different approaches, namely, MLR, PLS, hierarchical support
vector regression, SVM, radial basis function NN, and genetic
function algorithm. The RMSEs (log units) found for the 97-com-
pound training set were, respectively, 1.46, 1.66, 0.51, 0.71, 0.85,
and 0.90; for the 25-compound test set the RMSEs were 1.13,
0.91, 0.65, 0.85, 0.78, and 0.83 (OECD compliance 1,4IE).
Again, the NN approach did not yield the best results.
4.2.5 Acute Toxicity Tsakovska et al. [84] have reviewed the QSAR prediction of mam-
(Mammalian) malian toxicity. Devillers [85] employed PLS regression and back-
propagation NN to model acute oral toxicity (LD50) to the rat of
51 organophosphorus pesticides, with the use of physicochemical
descriptors. The NN approach yielded RMSE = 0.29 log unit for
the training set and 0.26 log unit for the test set (OECD compli-
ance 1,4IE); these figures were greatly superior to those from PLS
regression.
The oral toxicity of 54 benzodiazepine drugs to the mouse was
modeled by MLR, back-propagation ANN, SVM, and PLS meth-
ods [86]. RMSE values (log units) were, respectively, training set
0.213, 0.142, 0.183, and 0.174 and test set 0.194, 0.124, 0.192,
and 0.165, indicating the superior predictivity of the ANN method
(OECD compliance 1,4IE).
No mammalian inhalation toxicity QSAR studies appear to
have been done using a neural network approach.
4.2.6 Aquatic Toxicity Kaiser and Niculescu [87] used a probabilistic NN to model a large
data set of toxicity of organic chemicals to Daphnia magna, using a
combination of physicochemical and structural descriptors. For the
700-compound training set, s = 0.512 log unit, and for the 76-com-
pound test set s = 0.668 log unit (OECD compliance 1,4IE).
Another aquatic invertebrate that has been widely used in
QSAR toxicity studies is the ciliate Tetrahymena pyriformis. Over
2,000 chemicals have been tested in the same laboratory against
this organism, which means that experimental error should be
very low. Kahn et al. [88] compared the performance of MLR
and back-propagation NN on a large number of organic chemi-
cals, using CODESSA descriptors. The BPNN results were better
(914-compound training set s = 0.442 log unit, 457-compound
test set s = 0.484 log unit) than were those from MLR (training
set s = 0.551 log unit, test set s = 0.561 log unit) (OECD compli-
ance 1,4IE). Kahn et al. also reported 18 previous ANN QSAR
studies on toxicity to T. pyriformis.
Izadiyan et al. [89] modeled the effects of 40 ionic chemicals
on the limnic green alga Scenedesmus vacuolatus, using both MLR
and multilayer perceptron NN, with structural descriptors.
Standard errors of prediction in the validation set were MLR 0.525
log unit and MLPNN 0.440 log unit (OECD compliance 1,4IE).
4.2.7 Degradation Pesticide field half-lives were modeled [90] in a classification

(Ready Biodegradability) approach with a back-propagation NN, using structural fragments
as descriptors. For 110 training-set compounds and 13 test-set
compounds, correct classifications were 95.5 % and 84.6 %, respec-
tively (OECD compliance 1,4IE). These results were much better
than those obtained with discriminant factor analysis.
Aerobic biodegradation of 21 aromatic chemicals was mod-
eled [91] by MLR, principal component regression, and back-
propagation NN, using quantum-mechanical descriptors. The
BPANN model yielded RMSE = 0.136 log unit (training set)
and 0.024 log unit (validation set), markedly better than the
results from the MLR and PCR models (OECD compliance
1,4IE,5).
4.2.8 Acute Dermal There do not appear to be any neural network QSAR studies of
Toxicity (Mammalian) mammalian toxicity through dermal absorption.
4.2.9 Short-Term A QSAR (No. Q17-10-31-264) in the (Q)SAR Model Reporting

Repeat-Dose Toxicity Format (QMRF) [28] uses a back-propagation NN to model
repeat-dose reproductive/developmental toxicity in the rat. The
reported statistics are RMSE (training set) = 0.434 log unit and
RMSE (test set) = 2.521 log unit (OECD compliance 1,3,4IE,5).
The fact that the test-set compounds are so poorly predicted sug-
gests either overtraining or that the test-set compounds were not
within the applicability domain of the training set.
4.2.10 Reproductive Most QSAR studies of reproductive toxicity have been made on
Toxicity endocrine disruption. Liu et al. [92] used least-square SVM, coun-
terpropagation NN, and k-nearest neighbor (k-NN) approaches to
classify the endocrine-disruption capability of 232 training-set
chemicals and 87 external test-set chemicals, using molecular
structure descriptors. The three methods gave very similar results,
with correct predictions as follows: LS-SVM 89.66 %, CP-NN
87.50 %, and k-NN 89.22 % for the training set and LS-SVM
83.91 %, CP-NN 82.76 %, and k-NN 83.91 % for the external test
set (OECD compliance 1,3,4IE).
Roncaglioni et al. [93] investigated the performance of five
different nonlinear QSAR models (decision forest (DF), adaptive
fuzzy partition (AFP), decision tree (CART), multilayer percep-
tron feed-forward NN (MLPNN), and support vector machine
(SVM)) to classify the endocrine-disruption capability of 232
training-set chemicals and 87 external test-set chemicals, using
topological and structural descriptors. The correct classifications
were DF 97.04 %, AFP 86.36 %, CART 85.38 %, MLPNN 84.39 %,
and SVM 89.92 % for the training set and DF 85.33 %, AFP
85.33 %, CART 85.33 %, MLPNN 84.00 %, and SVM 86.67 % for
the external test set (OECD compliance 1,3,4IE).
4.2.11 Toxicokinetics There are numerous toxicokinetic parameters, and many QSAR
studies have been made of most of them. Wajima et al. [94] inves-
tigated the prediction of human clearance from data for 68 drugs
on humans, dogs, and rats, with physicochemical descriptors and
dog and rat data, using several approaches. RMSEs (log units)
were MLR 0.350, PLS 0.350, and feed-forward NN 0.395 (OECD
compliance 1,4). No external test set was used.
Blood levels of the antibiotic tobramycin in about 300 patients
were modeled with a back-propagation NN, using patient data
(e.g., age, weight, sex, illness) as descriptors [95]. The MAE was
33.9 %, which was considered acceptable (OECD compliance 1,4).
No external test set was used. The authors concluded that “neural
networks were capable of capturing the relationships between
plasma drug levels and patient-related prognostic factors from rou-
tinely collected sparse within-patient pharmacokinetic data.”
4.2.12 Short-Term Fish Singh et al. [96] compared the performance of a number of artifi-
Toxicity cial intelligence approaches (multilayer perceptron NN, radial basis
function NN, generalized regression NN, gene expression program-
ming (GEP), SVM, and decision tree (DT)) to model the acute
fish toxicity of 573 chemicals, of which 458 formed the training
set and 115 the test set. The MAEs (log units) obtained for the
training set were MLPNN 0.54, RBFNN 0.57, GRNN 0.34, GEP
0.71, and DT 0.43, and those for the test set were MLPNN 0.46,
RBFNN 0.41, GRNN 0.37, SVM 0.49, GEP 0.48, and DT 0.54
(OECD compliance 1,4IE,5). The generalized regression NN in
particular performed very well.
Gong et al. [97] used radial basis function NN and a heuristic
method to model the acute fish toxicity of 92 substituted benzenes
(training set 74, test set 18 chemicals) using CODESSA descrip-
tors. The RMSEs (log units) for the training and test sets, respec-
tively, were RBFNN 0.220 and 0.205 and heuristic method 0.273
and 0.223 (OECD compliance 1,4IE,5).
4.2.13 Activated Sludge Okey and Martis [98] employed a feed-forward NN to model the
Respiration Inhibition inhibition of activated sludge respiration by 139 (training set) and
25 (test set) diverse organic chemicals, using quantum-mechanical
and topological descriptors. They obtained MAE = 0.41 log unit
for the test set, which is a reasonable result (OECD compliance
1,4E,5).
A QSAR (No. Q17-10-1-226) in the (Q)SAR Model Reporting
Format (QMRF) [28] employs a multilayer perceptron back-
propagation NN to model the inhibition of activated sludge respi-
ration, using physicochemical and structural descriptors. The
reported statistics are RMSE (68-chemical training set) = 0.029 log
unit and RMSE (15-chemical test set) = 0.03 log unit (OECD
compliance 1,2,3,4IE,5).
4.2.14 Abiotic The acid hydrolysis of a large number of carboxylic acid esters in
Degradation (Hydrolysis) water and in water–organic solvent mixtures was modeled [99]
using a feed-forward NN and quantum-chemical descriptors,
descriptors representing the various organic solvents, and temper-
ature and solvent concentration. For the training set (1,883
records), RMSE = 0.271 log unit, and for the test set (209 records),
RMSE = 0.342 log unit (OECD compliance 1,4IE).
4.2.15 Adsorption/ See Subheading 4.1.12 above.

Desorption
4.2.16 Sub-chronic Sub-chronic toxicity is defined as the ability of a toxic substance to

Rodent Toxicity cause toxic effects for more than 1 year but less than the lifetime of
the exposed organism. It thus excludes carcinogenicity. A recent
paper [100] makes it clear that sub-chronic toxicology is still a
young science, with only a few published prediction studies, and
apparently none using ANN approaches.
4.2.17 Prenatal There do not appear to be any neural network QSAR studies of
Developmental Toxicity either of these endpoints.
4.2.18 Two-Generation
Reproductive Toxicity
4.2.19 Long-Term Meng and Lin [101] used feed-forward back-propagation NNs to
Aquatic Toxicity model both acute and chronic (up to 56-day) toxicity of alcohol
ethoxylates (nonionic surfactants) to fathead minnow (Pimephales
promelas), Daphnia magna, and green alga (Pseudokirchneriella
subcapitata). Unfortunately they presented their results largely in
graphical form, making it virtually impossible to obtain accurate
numerical predictions, even though they used both training and
external test sets. It appears from the graphs that chronic toxicity
ran more or less in parallel to acute toxicity. For example, for D.
magna the mean ratio of acute median effect concentration to
chronic no-observed-effect concentration was 2.8.
4.2.20 Bioaccumulation Bioaccumulation incorporates both bioconcentration factor, BCF

in Aquatic Species (which relates to the uptake of chemicals by fish or other aquatic
species from water), and biomagnification, which is the increase in
concentration from one link in a food chain to another. For
REACH purposes, bioaccumulation is taken to mean BCF.
Fatemi et al. [102] used a genetic algorithm-feed-forward
back-propagation NN on a training set of fish BCF values for 44
diverse organic chemicals and a test set of 9 similar chemicals, using
topological and physicochemical descriptors. They obtained
s = 0.259 log unit for the training set and s = 0.398 log unit for the
test set (OECD compliance 1,4IE).
Zhao et al. [103] used MLR, radial basis function NN, and
SVM to model a large data set of fish BCF values, using topological
and physicochemical descriptors. For the 378-chemical training
set, standard errors of prediction (log units) were MLR 0.70,
RBFNN 0.58, and SVM 0.63, and for the 95-chemical test set, the
errors were MLR 0.74, RBFNN 0.63, and SVM 0.62. A hybrid
RBFNN, from two models with different descriptors, yielded
s = 0.56 for the training set and 0.59 for the test set (OECD com-
pliance 1,4IE).
4.2.21 Effects Honeybees are susceptible to pesticides, and Devillers et al. [104]
on Terrestrial Organisms developed a feed-forward back-propagation NN model of the
(Invertebrates, acute toxicity of 100 pesticides to Apis mellifera. Using the auto-
Microorganisms, Plants) correlation vectors of physicochemical properties as descriptors,
they obtained RMSE = 0.430 log unit for the 89-chemical training
set and 0.386 log unit for the 11-chemical test set (OECD compli-
ance 1,4IE). These good results contrasted sharply with poor
results from a PLS model (data not given).
A study of fungicidal activities of thiazoline derivatives against
rice blast (Magnaporthe grisea) used both MLR and feed-forward
back-propagation NN, with physicochemical descriptors [105].
Three sets of thiazoline derivatives, with different substitution pat-
terns, were used, and similar good results were obtained for all
three sets. The largest set used 82 training compounds (standard
error (log units) = 0.139 (MLR) and 0.097 (NN)) and 18 test
compounds (standard error (log units) = 0.162 (MLR) and 0.122
(NN)) (OECD compliance 1,4IE).
4.2.22 Carcinogenicity It is perhaps surprising that carcinogenicity determination is

required only for production rates over 1,000 tonnes/year.
Counterpropagation NN was used by Fjodorova et al. [106]
to model both qualitative and quantitative measures of rat carcino-
genicity, using topological and molecular properties. For a
644-compound training set and 27 descriptors, they obtained
RMSE = 1.52 log unit, while for a 161-chemical test set, they
obtained RMSE = 1.80 (OECD compliance 1,4E). Strangely, this
model was developed using both carcinogens and noncarcinogens.
However, when they redeveloped the model using only the 421
carcinogens, the model was no better. The authors comment that
their results for the test set “are not high enough to fulfill criteria
for the predictive power of QSAR models concerning prediction of
carcinogenic potency.” They therefore developed a qualitative
(classification) model using the same training and test sets as
before. For the training set, they obtained 92.2 % correct classifica-
tion but only 68.3 % for the test set (OECD compliance 1,4E).
Singh et al. [107] also employed NN approaches to model
both qualitative and quantitative measures of carcinogenicity of
rodents, using two NN methods with physicochemical, structural,
and topological descriptors. For the qualitative (classification)

work on rat data, they used probabilistic NN and, with a training
set of 373 active and 294 inactive chemicals, found 93.85 % correct
classification; for the test set of 93 active and 74 inactive chemicals,
they found 85.03 % correct classification (OECD compliance
1,2,4IE). For their quantitative modeling, they employed general-
ized regression NN on 457 carcinogens tested in the rat and, with
an unspecified number of training and test-set chemicals, found
MAE (training) = 0.44 log unit and MAE (test) = 0.72 log unit.
They then used the model to predict the carcinogenicity of 292
chemicals in the mouse (MAE = 0.68 log unit) and of 38 chemicals
in the hamster (MAE = 0.57 log unit) (OECD compliance 1,4IE).
4.2.23 Long-Term There do not appear to be any neural network QSAR studies of
Effects on Terrestrial long-term effects on terrestrial organisms.
Organisms (Invertebrates,
Microorganisms,
Plants, Birds)
4.3 Software Many computer packages are available for the prediction of physi-
for Property cochemical and toxicological properties. Some of these are freely
and Toxicity Prediction distributed; others are commercial. Dearden et al. [31] have
reviewed the software available for physicochemical property pre-
diction, and Dearden [108] and Lapenna et al. [109] have reviewed
the software available for toxicity prediction. Unfortunately there
is a general lack of transparency concerning the nature of the algo-
rithms used in these packages, and so it is impossible to say with
confidence, in many cases, whether or not the package uses neural
networks for prediction and whether there is OECD compliance.
Among those that predict physicochemical properties, some
are clearly described as using methods other than neural networks.
For example, Molecular Modelling Pro [110] is described as mod-
eling melting and boiling points by the method of Joback and Reid
[111], which is a multiple linear regression method. Others use
neural network methods. For example, ADMET Predictor [112]
uses neural networks in the prediction of pKa, aqueous solubility,
and octanol–water partition coefficient. ChemSilico [113] also
uses neural network methods in the prediction of aqueous solubility
and octanol–water partition coefficient.
There is a similar situation in regard to toxicity prediction.
TOPKAT [114] is described by Lapenna et al. [109] as using
regression analysis to predict rat oral LD50. In contrast, TerraQSAR
[115] is described [116] as using neural net technology to predict
mouse and rat oral and intravenous LD50, fathead minnow and
Daphnia magna LC50, and skin irritation. ADMET Predictor
[112] uses neural networks to predict mutagenicity and endocrine
disruption, and ChemSilico [113] uses neural networks to predict
mutagenicity.
5 Conclusions
Various types of artificial neural networks are now widely used in

the prediction of physicochemical properties and a wide range of
toxicity endpoints relevant to REACH requirements. Generally
ANNs perform well in both classification and quantitative end-
point models. The ANN QSAR models reported herein are accept-
able with regard to some of the OECD Guidelines for the Validation
of (Q)SARs, usually Guidelines 1 and 4. However, there is no rea-
son why ANN QSAR models cannot comply with Guideline 3 as
well and perhaps with Guideline 5 also. It is unfortunate that, to
date, very few commercially available software packages for prop-
erty and toxicity prediction give any indication on their websites as
to whether or not ANNs are used in their predictive models.
Acknowledgments
We are grateful to Dr. T.I. Netzeva and Dr. A.P. Worth for valuable

comments on the draft manuscript.
References
1. European Parliament: Regulation (EC) N° 8. ECHA Guidance on information require-

1907/2006 of the European Parliament ments and safety assessment. Chapter R.6.
and of the Council of 18 December 2006 QSARs and grouping of chemicals, 2008.
concerning the Registration, Evaluation, Available at http://echa.europa.eu/docu-
Authorisation and Restriction of Chemicals ments/10162/13632/information_require-
(REACH). Available at http://eur-lex. ments_r6_en.pdf. Accessed 5 Jan 2014
europa.eu/LexUriServ/LexUriServ.do?uri= 9. OECD Principles for the Validation of (Q)
oj:l:2006:396:0001:0849:en:pdf. Accessed SARs. Available at http://www.oecd.org/
5 Jan 2014 dataoecd/33/37/37849783.pdf. Accessed 5
2. Hansch C, Clayton JM (1973) Lipophilic Jan 2014
character and activity of drugs II: the para- 10. OECD Environment Directorate, Joint
bolic case. J Pharm Sci 62:1–23 Meeting of the Chemicals Committee and the
3. Cronin MTD, Gregory BW, Schultz TW Working Party on Chemicals, Pesticides and
(1998) Quantitative structure-activity analy- Biotechnology. Available at http://www.olis.
ses of nitrobenzene toxicity to Tetrahymena oecd.org/olis/2004doc.nsf/LinkTo/
pyriformis. Chem Res Toxicol 11:902–908 NT00009192/$FILE/JT00176183.PDF.
4. Devillers J (ed) (1996) Neural networks in Accessed 5 Jan 2014
QSAR and drug design. Academic, London 11. Fjodorova N, Novich M, Vrachko M et al
5. Gasteiger J, Engel T (2003) (2008) Directions in QSAR modeling for
Chemoinformatics—a textbook. Wiley, regulatory use in OECD member countries,
Weinheim EU and in Russia. J Environ Sci Health C
6. Witten IH, Frank E (2005) Data mining: Environ Carcinog Ecotoxicol Rev 26:
practical machine learning tools and tech- 201–236
niques. Morgan Kaufmann, San Francisco 12. Devillers J (2005) A new strategy for using
7. Devillers J (1996) Strengths and weaknesses supervised artificial neural networks in
of the backpropagation neural network QSPR QSAR. SAR QSAR Environ Res 16:433–442
studies. In: Devillers J (ed) Neural networks 13. Baskin II, Palyulin VA, Zefirov NS (2008)
in QSAR and drug design. Academic, Neural networks in building QSAR models.
London, pp 1–46 In: Livingstone DS (ed) Artificial neural
networks: methods and protocols. Humana 26. Vračko M, Bandelj V, Barbier P et al (2006)
Press, New York, pp 137–158 Validation of counter propagation neural
14. Devillers J (2009) Artificial neural network network models for predictive toxicology
modeling of the environmental fate and eco- according to the OECD principles: a case
toxicity of chemicals. In: Devillers J (ed) study. SAR QSAR Environ Res 17:265–284
Ecotoxicology modeling. Springer, Dordrecht, 27. Worth AP (2010) The role of QSAR method-
pp 1–28 ology in the regulatory assessment of chemi-
15. Gleeson MP, Modi S, Bender A et al (2012) cals. In: Puzyn T, Leszczynski J, Cronin MTD
The challenges involved in modelling toxicity (eds) Recent advances in QSAR studies: meth-
data in silico: a review. Curr Pharm Des 18: ods and applications. Springer, Dordrecht,
1266–1291 pp 367–382
16. Moore DRJ, Breton RL, MacDonald DB 28. (Q)SAR Model Reporting Format. Available
(2003) A comparison of model performance at http://qsardb.jrc.it/qmrf. Accessed 16
for six quantitative structure-activity relation- Jan 2014
ship packages that predict acute toxicity to 29. ECHA Guidance on information require-
fish. Environ Toxicol Chem 22:1799–1809 ments and chemical safety assessment.
17. Fatemi MH, Abraham MH, Haghdadi M Chapter R.7a. Endpoint specific guidance, in
(2009) Prediction of biomagnification factors Guidance for the Implementation of REACH,
for some organochlorine compounds using ECHA, Helsinki, 2012. Available at http://
linear free energy relationship parameters and echa.europa.eu/support/guidance-on-reach-
artificial neural networks. SAR QSAR Environ and-clp-implementation/consultation-
Res 20:453–465 procedure. Accessed 25 Jan 2014
18. Tan N-X, Li P, Rao H-B et al (2010) 30. Regulation (EC) No. 1907/2006 of the
Prediction of the acute toxicity of chemical European Parliament and of the Council of
compounds to the fathead minnow by 18 December 2006. Available at http://
machine learning approaches. Chemometr eurlex.europa.eu/LexUriServ/LexUriServ.
Intell Lab Syst 100:66–73 do?uri=CONSLEG:2006. Accessed 25 Jan
19. Dearden JC, Cronin MTD, Kaiser KLE 2014
(2009) How not to develop a quantitative 31. Dearden JC, Rotureau P, Fayet G (2013)
structure-activity or structure-property rela- QSPR prediction of physico-chemical proper-
tionship (QSAR/QSPR). SAR QSAR ties for REACH. SAR QSAR Environ Res
Environ Res 20:241–266 24:279–318
20. Baskin II, Ait AO, Halberstam NM et al 32. Taskinen J, Yliruusi J (2003) Prediction of
(2002) An approach to the interpretation of physicochemical properties based on neural
backpropagation neural network models in network modelling. Adv Drug Deliv Rev
QSAR studies. SAR QSAR Environ Res 13: 55:1163–1183
35–41 33. Godavarthy SS, Robinson RL, Gasem KAM
21. Guha R, Jurs PC (2005) Interpreting com- (2006) An improved structure-property model
putational neural network QSAR models: a for predicting melting-point temperatures. Ind
measure of descriptor importance. J Chem Eng Chem Res 45:5117–5126
Inf Model 45:800–806 34. Karthikeyan M, Glen RC, Bender A (2005)
22. Guha R, Stanton DT, Jurs PC (2005) General melting point prediction based on a
Interpreting computational neural network diverse compound data set and artificial neu-
QSAR models: a detailed interpretation of the ral networks. J Chem Inf Model 45:581–590
weights and biases. J Chem Inf Model 45: 35. Hall LH, Story CT (1996) Boiling point and
1109–1121 critical temperature of a heterogeneous data
23. Chaudhry Q, Piclin N, Cotterill J et al (2010) set. QSAR with atom type electrotopological
Global QSAR models of skin sensitisers for state indices using artificial neural networks.
regulatory purposes. Chem Cent J 4(Suppl 1): J Chem Inf Comput Sci 36:1004–1014
S1–S6 36. Chalk AJ, Beck B, Clark T (2001) A quantum
24. Johnson SR (2008) The trouble with QSAR mechanical/neural network model for boiling
(or how I learned to stop worrying and points with error estimation. J Chem Inf
embrace fallacy). J Chem Inf Model 48:25–26 Comput Sci 41:457–462
25. Enoch SJ, Madden JC, Cronin MTD (2008) 37. Gakh AA, Gakh EG, Sumpter BG et al (1994)
Identification of mechanisms of toxic action Neural network-graph theory approach to the
for skin sensitisation using a SMARTS pattern prediction of the physical properties of
based approach. SAR QSAR Environ Res organic compounds. J Chem Inf Comput Sci
19:555–578 34:832–839
38. Valderrama JO, Reátegui A, Rojas RE (2009) 51. Kier LB, Hall LH (1999) Molecular structure
Density of ionic liquids using group contribu- description: the electrotopological state.
tion and artificial neural networks. Ind Eng Academic, New York
Chem Res 48:3254–3259 52. Katritzky AR, Stoyanova-Slavova IB, Dobchev
39. McClelland HE, Jurs PC (2000) Quantitative DA et al (2007) QSPR modelling of flash
structure-property relationships for the pre- points: an update. J Mol Graph Model 26:
diction of vapor pressures of organic com- 529–536
pounds from molecular structure. J Chem Inf 53. Gharagheizi F, Alamdari RF, Angaji MT
Comput Sci 40:967–975 (2008) A new neural network-group contri-
40. Chalk AJ, Beck B, Clark T (2001) A bution method for estimation of flash point
temperature-dependent quantum mechanical/ temperature of pure components. Energy
neural net model for vapor pressure. J Chem Fuel 22:1628–1635
Inf Comput Sci 41:1053–1059 54. Gharagheizi F (2010) Chemical structure-
41. Roosta A, Setoodeh P, Jahanmiri A (2012) based model for estimation of the upper flam-
Artificial neural network modeling of surface mability limit of pure compounds. Energy
tension for pure organic compounds. Ind Eng Fuel 24:3867–3871
Chem Res 51:561–566 55. Gharagheizi F (2009) Prediction of upper
42. Gharagheizi F, Eslamimanesh A, Mohammadi flammability limit percent of pure compounds
AH et al (2011) Use of artificial neural from their molecular structures. J Hazard
network- group contribution method to Mater 167:507–510
determine surface tension of pure com- 56. Gharagheizi F (2008) Quantitative
pounds. J Chem Eng Data 56:2587–2601 structure-property relationship for prediction
43. Dearden JC (2006) In silico prediction of of the lower flammability limit of pure com-
aqueous solubility. Expert Opin Drug Discov pounds. Energy Fuel 22:3037–3039
1:31–52 57. Gharagheizi F (2009) A new group
44. Yan A, Gasteiger J (2003) Prediction of aque- contribution- based model for estimation of
ous solubility of organic compounds based on lower flammability limit of pure compounds.
a 3D structure representation. J Chem Inf J Hazard Mater 170:595–604
Comput Sci 43:429–434 58. Cho SG, No KT, Goh EM et al (2005)
45. Katritzky AR, Wang Y, Sild S et al (1998) Optimization of neural networks architecture
QSPR studies on vapor pressure, aqueous for impact sensitivity of energetic molecules.
solubility, and the prediction of water-air par- Bull Korean Chem Soc 26:399–408
tition coefficients. J Chem Inf Comput Sci 59. Wang R, Jiang J, Pan Y et al (2009)
38:720–725 Prediction of impact sensitivity of nitro ener-
46. Bruneau P (2001) Search for predictive getic compounds by neural network based
generic model of aqueous solubility using on electrotopological-state indices. J Hazard
Bayesian neural nets. J Chem Inf Comput Sci Mater 166:155–186
41:1605–1616 60. Tetteh J, Metcalfe E, Howells SL (1996)
47. Livingstone DJ (2003) Theoretical property Optimisation of radial basis and backpropaga-
predictions. Curr Top Med Chem 3: tion neural networks for modelling auto-
1171–1192 ignition temperature by quantitative
48. Klopman G, Zhu H (2005) Recent method- structure-property relationships. Chemometr
ologies for the estimation of n-octanol-water Intell Lab Syst 32:177–191
partition coefficients and their use in the pre- 61. Pan Y, Jiang J, Wang R et al (2008) Prediction
diction of membrane transport properties of of auto-ignition temperatures of hydrocar-
drugs. Mini Rev Med Chem 5:127–133 bons by neural network based on atom-type
49. Chen H-F (2009) In silico log P prediction electrotopological-state indices. J Hazard
for a large data set with support vector Mater 157:510–517
machines, radial basis neural networks and 62. Liu GS, Yu JG (2005) QSAR analysis of soil
multiple linear regression. Chem Biol Drug sorption coefficients for polar organic chemi-
Des 74:142–147 cals: substituted anilines and phenols. Water
50. Tetko IV, Tanchuk VY, Villa AEP (2001) Res 39:2048–2055
Prediction of n-octanol-water partition 63. Goudarzi N, Goodarzi M, Araujo MCU et al
coefficients from PHYSPROP database (2009) QSPR modeling of soil sorption coef-
using artificial neural networks and E-state ficients (Koc) of pesticides using SPA-ANN
indices. J Chem Inf Comput Sci 41: and SPA-MLR. J Agric Food Chem 57:
1407–1421 7153–7158
64. Luan F, Ma W, Zhang H et al (2005) 76. Patlewicz GY, Rodford RA, Ellis G et al (2000)
Prediction of pKa for neutral and basic drugs A QSAR model for the eye irritation of cat-
based on radial basis function neural networks ionic surfactants. Toxicol In Vitro 14:79–84
and the heuristic method. Pharm Res 22: 77. Patlewicz G, Roberts DW, Uriarte E (2008)
1454–1460 A minireview of available skin sensitization
65. CODESSA software. Available at www.semi- (Q)SARs/expert systems. Chem Res Toxicol
chem.com. Accessed 26 Jan 2014 21:521–541
66. Habibi-Yangjeh A, Danandeh-Jenagharad M, 78. Devillers J (2000) A neural network SAR
Nooshyar M (2005) Prediction acidity constant model for allergic contact dermatitis. Toxicol
of various benzoic acids and phenols in water Methods 10:181–193
using linear and nonlinear QSPR models. Bull 79. Cronin MTD, Basketter DA (1994)
Korean Chem Soc 26:2007–2016 Multivariate QSAR analysis of a skin sensiti-
67. Kauffman GW, Jurs PC (2001) Prediction of zation database. SAR QSAR Environ Res 2:
surface tension, viscosity and thermal conduc- 159–179
tivity for common organic solvents using 80. Golla S, Madihally S, Robinson RL et al
quantitative structure-property relationships. (2009) Quantitative structure-property rela-
J Chem Inf Comput Sci 41:408–418 tionship modeling of skin sensitization: a
68. Artemenko NV, Baskin II, Palyulin VA et al quantitative prediction. Toxicol In Vitro 23:
(2001) Prediction of physical properties of 454–465
organic compounds using artificial neural net- 81. Benigni R (ed) (2003) Quantitative structure-
works within the substructure approach. activity relationship (QSAR) models of muta-
Doklady Chem 381:317–320 gens and carcinogens. CRC Press, Boca Raton
69. Modarresi H, Modarress H, Dearden JC 82. Xu C, Cheng F, Chen L et al (2012) In silico
(2007) QSPR model of Henry’s law constant prediction of chemical Ames mutagenicity.
for a diverse set of organic chemicals based on J Chem Inf Model 52:2840–2847
genetic algorithm-radial basis function net- 83. Leong MK, Lin S-W, Chen H-B et al (2010)
work approach. Chemosphere 66:2067–2076 Predicting mutagenicity of aromatic amines
70. Gharagheizi F, Abbasi R, Tirandazi B (2010) by various machine learning approaches.
Prediction of Henry’s law constant of organic Toxicol Sci 116:498–513
compounds in water from a new group- 84. Tsakovska I, Lessigiarska I, Netzeva T et al
contribution-based method. Ind Eng Chem (2008) A mini review of mammalian toxicity
Res 49:10149–10152 (Q)SAR models. QSAR Comb Sci 27:41–48
71. ECHA Guidance on information require- 85. Devillers J (2004) Prediction of mammalian
ments and chemical safety assessment. toxicity of organophosphorus pesticides from
Chapter R.7b. Endpoint specific guidance, in QSTR modelling. SAR QSAR Environ Res
Guidance for the Implementation of REACH, 15:501–510
ECHA, Helsinki, 2012. Available at http://
echa.europa.eu/support/guidance-on-reach- 86. Funar-Timofei S, Ionescu D, Suzuki T (2010)
and-clp-implementation/consultation- A tentative quantitative structure-toxicity
procedure. Accessed 17 Jan 2014 relationship study of benzodiazepine drugs.
Toxicol In Vitro 24:184–200
72. Salina AG, Patlewicz G, Worth AP (2008) A
review of QSAR models for skin and eye irri- 87. Kaiser KLE, Niculescu SP (2001) Modeling
tation and corrosion. QSAR Comb Sci 27: acute toxicity of chemicals to Daphnia magna:
49–59 a probabilistic neural network approach.
Environ Toxicol Chem 20:420–431
73. Golla S, Madihally S, Robinson RL et al
(2009) Quantitative structure-property rela- 88. Kahn I, Sild S, Maran U (2007) Modeling the
tionships modeling of skin irritation. Toxicol toxicity of chemicals to Tetrahymena pyrifor-
In Vitro 23:176–184 mis using heuristic multilinear regression and
heuristic back-propagation neural networks.
74. Barratt MD (1996) Quantitative structure- J Chem Inf Model 47:2271–2279
activity relationships (QSARs) for skin corro-
sivity of organic acids, bases and phenols: 89. Izadiyan P, Fatemi MH, Izadiyan M (2013)
principal components and neural networks Elicitation of the most important structural
analysis of extended datasets. Toxicol In Vitro properties of ionic liquids affecting ecotoxic-
10:85–94 ity in limnic green algae: a QSAR approach.
Ecotoxicol Environ Saf 87:42–48
75. Barratt MD (1997) QSARs for the eye irrita-
tion potential of neutral organic chemicals. 90. Domine D, Devillers J, Chastrette M et al
Toxicol In Vitro 11:1–8 (1993) Estimating pesticide field half-lives
from a backpropagation neural network. SAR 103. Zhao C, Boriani E, Chana A et al (2008) A
QSAR Environ Res 1:211–219 new hybrid system of QSAR models for pre-
91. Jing G-H, Li X-L, Zhou Z-M (2011) diction bioconcentration factors (BCF).
Quantitative structure-biodegradability rela- Chemosphere 73:1701–1707
tionship study about the aerobic biodegrada- 104. Devillers J, Pham-Delègue MH, Decourtye A
tion of some aromatic compounds. Chin J et al (2002) Structure-toxicity modelling of
Struct Chem 30:368–375 pesticides to honey bees. SAR QSAR Environ
92. Liu H, Papa E, Walker JD et al (2007) In Res 13:641–648
silico screening of estrogen-like chemicals 105. Song JS, Moon T, Nam KD et al (2008)

based on different nonlinear classification Quantitative structure-activity relationship
models. J Mol Graph Model 26:135–144 (QSAR) studies for fungicidal activities of
93. Roncaglioni A, Piclin N, Pintore M et al thiazoline derivatives against rice blast. Bioorg
(2008) Binary classification models for Med Chem Lett 18:2133–2142
endocrine disrupter effects mediated through 106. Fjodorova N, Vračko M, Tušar M et al (2010)
the estrogen receptor. SAR QSAR Environ Quantitative and qualitative models for carci-
Res 19:697–733 nogenicity prediction for non-congeneric
94. Wajima T, Fukumura K, Yano Y et al (2002) chemicals using CP NN method for regula-
Prediction of human clearance from animal tory uses. Mol Divers 14:581–594
data and molecular structural parameters 107. Singh KP, Gupta S, Rai P (2013) Predicting
using multivariate regression analysis. J Pharm carcinogenicity of diverse chemicals using
Sci 91:2489–2499 probabilistic neural network modelling
95. Chow H-H, Tolle KM, Roe DJ et al (1997) approaches. Toxicol Appl Pharmacol 272:
Application of neural networks to population 465–475
pharmacokinetic data analysis. J Pharm Sci 108. Dearden JC (2010) Expert systems for tox-
86:840–845 icity prediction. In: Cronin MTD, Madden
96. Singh KP, Gupta S, Rai P (2013) Predicting JC (eds) In silico toxicology: principles and
acute aqueous toxicity of structurally diverse applications. Royal Society of Chemistry,
chemicals in fish using artificial intelligence London, pp 478–507
approaches. Ecotoxicol Environ Saf 95: 109. Lapenna S, Fuart-Gatnick M, Worth A (2010)
221–232 Review of QSAR models and software tools
97. Gong Z, Xia B, Zhang R et al (2008) for predicting acute and chronic systemic tox-
Quantitative structure-activity relationship icity. Available at http://ihcp.jrc.ec.europa.
study on fish toxicity of substituted benzenes. eu/our_labs/predictive_toxicology/doc/
QSAR Comb Sci 27:967–976 EUR_24639_EN.pdf/view. Accessed 25 Jan
98. Okey RW, Martis MC (1999) Molecular level 2014
studies of the origin of toxicity: determination 110. Molecular Modeling Pro. Available at http://
of key variables and selection of descriptors. www.chemsw.com/Software-and-Solutions/
Chemosphere 38:1419–1427 L a b o r a t o r y - S o f t w a r e / D r a w i n g - a n d -
99. Halberstam NM, Baskin II, Palyulin VA et al Modeling-Tools/Molecular-Modeling-Pro.
(2002) Quantitative structure-conditions- aspx. Accessed 25 Jan 2014
property relationship studies. Neural network 111. Joback KG, Reid RC (1987) Estimation of
modelling of the acid hydrolysis of esters. pure-component properties from group-
Mendeleev Commun 12(5):185–186 contributions. Chem Eng Commun 57:
100. Dewhurst I, Renwick AG (2013) Evaluation 233–243
of the threshold of toxicological concern 112. ADMET Predictor. http://www.simulations-
(TTC)—challenges and approaches. Regul plus.com. Accessed 28 Jan 2014
Toxicol Pharmacol 65:168–177 113. ChemSilico. Available at www.chemsilico.

101. Meng Y, Lin B-L (2008) A feed-forward arti- com. Accessed 28 Jan 2014
ficial neural network for prediction of the 114. TOPKAT. Available at www.accelrys.com.

aquatic ecotoxicity of alcohol ethoxylate. Accessed 28 Jan 2014
Ecotoxicol Environ Saf 71:172–186 115. TerraBase. Available at www.terrabase-inc.

102. Fatemi MH, Jalali-Heravi M, Konuze E
com. Accessed 28 Jan 2014
(2003) Prediction of bioconcentration factor 116. Kaiser KLE (2003) Neural networks for effect
using genetic algorithm and artificial neural prediction in environmental health issues using
network. Anal Chim Acta 486:101–108 large datasets. QSAR Comb Sci 22:185–190
Chapter 6
Artificial Neural Network for Charge Prediction

in Metabolite Identification by Mass Spectrometry
J.H. Miller, B.T. Schrom, and L.J. Kangas
Abstract
Collision-induced dissociation (CID) is widely used in mass spectrometry to identify biologically important
molecules by gaining information about their internal structure. Interpretation of experimental CID spec-
tra always involves some form of in silico spectra of potential candidate molecules. Knowledge of how
charge is distributed among fragments is an important part of CID simulations that generate in silico spec-
tra from the chemical structure of the precursor ions entering the collision chamber. In this chapter we
describe a method to obtain this knowledge by machine learning.
Key words Mass spectrometry, Collision-induced dissociation, Charge transfer, Metabolite identifica-
tion, Lipid metabolites
1 Introduction
Mass spectrometry (MS) is an analytical method that measures the

mass-to-charge ratio (m/z) of charged particles. In some cases this
quantity alone is sufficient to identify the molecules in the sample
being analyzed. For biological samples that contain large mole-
cules, fragmentation patterns in addition to precursor-ion mass are
usually required for identification. Collision-induced dissociation
(CID) is a commonly used method of fragmenting ions to obtain
a fingerprint of fragment masses that is unique to the molecules
under investigation. Precursor ions are accelerated toward, and
collide with, neutral atoms in a gas such as helium. These collisions
transform some of the kinetic energy into internal energy, which
causes bonds to break by a mechanism similar to thermal dissocia-
tion. The fragment ions are then analyzed by the mass spectrome-
ter in the usual way. Patterns of fragmentation obtained by this
type of experiment are commonly referred to as MS/MS spectra.
Software tools like SEQUEST [1] and Mascot [2] are widely
used in the application of MS/MS spectra for protein identifica-
tion. Since the peptide bond between amino acids in a protein is
89
90 J.H. Miller et al.
the most likely bond to break in CID, in silico MS/MS spectra can
be derived from protein sequence without detailed knowledge of
protein structure. For metabolites, the connection between molec-
ular structure and CID is much more complex than for polypep-
tides; consequently, software tools for metabolite identification
from MS/MS spectra have been more difficult to develop [3].
MetISIS [4] is a software package being developed to aid in
the identification of metabolites. Given a MS/MS spectrum
acquired from an unknown sample, MetISIS attempts to match the
spectrum to an entry in a database of in silico spectra developed
from the chemical structure of metabolites. Most of the entries in
the database can be ignored because their mass is significantly dif-
ferent from the mass of the precursor ions fragmented to produce
the MS/MS spectrum of the unknown. If the masses of the
unknown metabolite and the database candidates are within a spec-
ified tolerance, then the similarity of CID spectra is quantified in
various ways, including Pearson correlation coefficient (PCC).
Requiring similarity of precursor-ion mass and PCC near unity
produces a shortlist of candidate molecules with a high probability
of being the unknown metabolite.
The success of this identification technique depends on the
quality of the database of MS/MS spectra. It should contain a
large number of metabolites for which high-resolution CID spec-
tra are available. Populating such a database with experimental
MS/MS spectra is both time-consuming and costly. To circumvent
this difficultly, in silico databases, such as LipidBlast [5], have been
developed. MetISIS [4] generates in silico MS/MS spectra of
metabolites based on their molecular structure, such as those found
in the LIPID Metabolites and Pathways Strategy (LIPID MAPS)
database [6] and illustrated in Fig. 1 for phosphatidylcholine
18:0/18:0 (see Note 1).
The simulation of CID that MetISIS [4] performs to generate
an MS/MS spectrum for a metabolite of known chemical structure
has two components: (1) prediction of which bonds are most likely
to break in a collision and (2) prediction of which fragments will
continue to carry charge and be detected by the mass spectrometer.
5 524.58 506.42 184

O 34
O
45 21 19 17 15 13 11 9 2 23 32
4 1 P 25 27
O O 29
46 O 28
22 20 18 16 14 12 10 8 3 24 HO +
53 49 47 39 37 26 N
55 51 43 41 33
O H
35 7 6 31
54 52 50 48 44 42 40 38 506.42
30
524.58
O
36
Fig. 1 Molecular structure of phosphatidylcholine 18:0/18:0. Arrows mark bond cleavages in collision-induced
dissociation (CID) that generate the fragments most frequently observed in a linear ion-trap mass
spectrometer
Charge Prediction in Metabolite Identification 91
MetISIS uses machine learning by artificial neural networks (ANN)

to make both types of prediction. The charge-prediction problem is
the main focus of this chapter.
Collision-induced dissociation (CID) is a slow fragmentation
process, which means that most collisions change the internal
energy of ions without producing fragmentation. Consequently,
ions have the opportunity to relax to a state where the location of
charge minimizes their free energy. When the precursor ion con-
tains an atom like nitrogen with a low ionization potential, charge
is likely to remain localized during the dissociation process and the
question of which fragment is charged reduces to which fragment
has the nitrogen atom. Dissociation events of this type are com-
monly referred to as “neutral loss.”
Three mechanisms of neutral loss are marked by the arrows in
Fig. 1. Due to the symmetry of phosphatidylcholine 18:0/18:0,
neutral loss of a 266 Da tail can occur by cleavage of C-O bonds
between C4 and O3 and between C36 and O7. Also due to sym-
metry, neutral loss of a 284 Da tail can occur by breaking of bonds
between C2 and O3 and between C1 and O7. These neutral losses
give rise to ions of mass 524 and 506 Da, respectively. The third
mechanism of neutral loss shown in Fig. 1 is loss of both tails by
cleavage of the bond between C23 and O24, which produces an
ion of 184 Da.
Production of the 184 Da ion by loss of both tails is by far the
most frequent event in the fragmentation of phosphatidylcholine
18:0/18:0. It is 93 times more frequent than loss of the 284 Da
tail and 116 times more frequent than loss of the 266 Da tail.
Cleavage of the bond between C23 and O24 can cause charge
transfer, resulting in a zwitterion (an overall neutral fragment) of
mass 183 Da and an ion of mass 608 Da, but the neutral loss of
both tails by the same bond cleavage is almost 3,000 times more
frequent. Mechanisms of neutral loss are more complex in lipids
that contain more than one nitrogen atom, and in some cases, the
relative stability of positively charged atoms is not clear from the
molecular structure. In future research, computational chemistry
methods will be used to investigate the relative stability of sites of
positive charge.
A schematic of the CID simulation and associated machine-
learning algorithm developed by Lars Kangas as part of his doc-
toral thesis research is shown in Fig. 2.
A Monte Carlo simulation of the CID process is performed on a
large number of identical precursor ions of known molecular structure.
The physical model of dissociation induced by ion-atom collisions
must be consistent with the characteristics of the mass spectrometer
used to acquire the MS/MS spectra for machine learning and met-
abolic analysis. Initially, we chose to simulate CID in a linear ion-
trap mass spectrometer because the acceleration of precursor ions in
that type of spectrometer is a resonance process (see Note 2). Fragments
Fig. 2 Flow chart of the algorithm developed to learn the relative frequency of
bond cleavages in collision-induced dissociation (CID) from experimental data.
When training is complete, the three components above the dotted line are
ignored and in silico CID spectra can be generated from the molecular structure
of lipid metabolites. The charge-prediction component is a separate machine-
learning task using the methods described in this chapter. (Reproduced from [4]
with permission from Oxford University Press)
of precursor-ion dissociation are off resonance and unlikely to

undergo secondary fragmentation, which allowed us to focus on
the chemistry of the precursor ion (i.e., how strong its bonds are
and where charge is likely to be located).
Since neutral fragments are not detected, an algorithm to gen-
erate in silico MS/MS spectra must have a component to predict
which fragment after a collision-induced dissociation continues to
carry the charge of the precursor ion. This charge-prediction com-
ponent is separate from the neural network used to learn the rela-
tive frequency of bond cleavages. The method section of this
chapter describes a machine-learning approach to charge predic-
tion that uses the same input vectors as the ANN used to predict
frequencies of bond cleavage.
In the machine-learning phase, weights of the neural network
component in Fig. 2 are optimized through a genetic algorithm
[7]. The exemplar database for this phase of machine learning was
derived from experimental CID spectra acquired on a linear ion-
trap spectrometer with pure samples of the type of metabolites
under investigation. Lipid metabolites were chosen to demonstrate
a proof of concept because their CID spectra were known to be
relatively simple. High-quality CID spectra were obtained for
about 100 molecular species distributed more or less uniformly
over the major classes of lipid metabolites (see Note 3).
Fig. 3 A block in an input vector displaying 18 options for how an atom is bonded to its successor in the tree
derived from of a breadth-first search of a fragment from its root atom
Both the position (m/z) and intensity of peaks in experimental

MS/MS spectra were used to find patterns of bond cleavage, giving
rise to detectable charged fragments of the precursor ion. Spectra
of singly charged fragments provide most of the data for pattern
discovery. After weights in the neural network component con-
verged, the process of generating in silico CID spectra for a large
database of metabolites of known chemical structure can begin.
The LIPID MAPS database [6] contains more than 22,000 such
structures. This application of the algorithm illustrated by Fig. 2
does not require the components above the dotted line.
A method to input a molecular structure into the Monte Carlo
simulation of CID is fundamental to the machine-learning process
and subsequent generation of in silico spectra by MetISIS. Faulon
et al. [8] showed that molecules could be represented by undi-
rected graphs of atoms and bonds stored in adjacency matrices.
Tree structures and vectors for machine learning can be derived
from adjacency matrices [9] by selecting one atom to be the root
and constructing the predecessor graph of a breadth-first search
[10] to a specified depth. MetISIS [4] uses vectors, made up of
blocks like that shown in Fig. 3, to encode the fragments that are
produced when a bond in the chemical structure of the precursor
ion is cleaved during CID.
Most lipid metabolites are made up of six atom types: carbon
(C), hydrogen (H), oxygen (O), nitrogen (N), sulfur (S), and
phosphorus (P). An encoding scheme that records how a given
atom is connected to another atom can be constructed with a block
that is a vector with 18 components. The first six components are
CHONSP for single-bond connections, the next six components
are for double-bond connections, and the final six components are
for triple-bond connections. The root atom of a fragment is the
atom involved in the bond that was cleaved to produce the frag-
ment. A bonding pattern that connects the root atom to any other
atom in the fragment results in a series of CHONSP components
between the selected atom and the root. Since all of the six atom
types and three bond types are possible, the upper bound on the
number of paths for a distance (in bonds), D, away from the frag-
mentation site is (6)(6 × 3)D−1. We find that D = 8 is adequate to
capture the chemical neighborhood of bonds in lipid metabolites
(see Note 4).
Fig. 4 The distribution of true- and false-positive identifications as a function of Pearson correlation coefficient
(R2) in a test of the MetISIS algorithm that did not contain the charge-prediction component (see Fig. 2). While
the algorithm performed reasonably well, the approximately 50 false positives with R2 near unity suggested
that charge prediction was essential for accurate simulation of MS/MS spectra
In cases where the major peaks in the MS/MS spectrum of a

metabolite are well separated and the spectrum also contains a peak
for the precursor ion, charge prediction in simulations of CID in a
linear ion-trap spectrometer is of minor importance. Accurate pre-
dictions of the frequency of bond cleavage can be used to generate
potential fragment masses for the major peaks in the MS/MS spec-
trum. Given a close alignment of a predicted fragment mass with a
peak in the experimental spectrum, the mass of the precursor ion
can be used to determine the mass of the corresponding neutral
fragment, thereby allowing its peak to be eliminated from the in
silico spectrum. Figure 4 shows the results of an early test of
MetISIS performance without charge prediction.
Without charge prediction, the number of false-positive iden-
tifications is significant even when the PCC is close to unity. This
undesirable result could be traced directly to the cases where simu-
lation of CID predicted masses of neutral fragments that were too
close to peaks in experimental spectra to distinguish them from
charged fragments. These results clearly showed a need to predict
which fragments are charged after CID events.
2 Methods
2.1 Construction 1. The construction of training and testing sets begins by acquiring
of Training and Testing experimental CID spectra from pure samples of specified
Sets metabolites. The training/testing process for charge prediction
in CID of lipid metabolites was based on MS/MS experiments
with 94 lipids of known chemical structure (see Note 3).
2. The molecular structure of each lipid metabolite with an
experimental CID spectrum was provided by the chemical
supplier.
3. Input vectors (see Note 4) for all bond cleavages that might
be associated with peaks in experimental spectra were gener-
ated from the molecular structures. Bonds to hydrogen
atoms for which cleavage would reduce the precursor-ion
mass by only 1 Da were ignored. Bonds between carbon atoms
in tails of lipids were also ignored because it seemed likely
that any experimental peak could be explained by this type of
fragmentation.
4. To generate the output label, the left and right mass fields in
the <meta data> block of the encoded vector (see Note 4) were
used to determine the mass for each fragment and the total
mass of the lipid. The experimental CID spectrum of the lipid
was then examined to see if it contained a peak at either of
these two masses. In the absence of such a peak, the potential
input vectors for a charged and a neutral fragment were
removed from consideration. If there was a peak in the experi-
mental spectrum at only one of potential fragment masses,
then the exemplar with that mass in left-hand fragment was
labeled 0.8 and the exemplar with that mass as right-hand frag-
ment was labeled 0.2. If peaks were found at both masses, then
the mass with the higher intensity was considered the winner
and the output of the pair of exemplars was labeled accord-
ingly. A tolerance of ±2 Da in matching peak masses with frag-
ment masses was used. For the 44 lipids selected to generate a
training set, this process gave rise to 658 pairs of exemplars.
For the 50 lipids selected to generate a testing set, the same
process gave rise to 2,183 pairs of exemplars.
5. Both training and testing sets contained labeled exemplars
from all 15 classes of lipids (see Note 3) for which experimental
MS/MS spectra were available.
2.2 ANN Structure As shown in Fig. 5, the ANN architecture had an input layer with
the 1,270 nodes required for input vectors that characterize the
chemical environment around bond-cleavage sites (see Note 4) in
the lipids with experimental spectra (see Note 3).
Fig. 5 Structure of the neural network used to develop charge prediction for a set
of 94 lipid metabolites distributed over 15 classes (see Note 3). The input vector
size (1,270) reflects the amount of compression that was possible with this set
of molecules (see Note 4). Bias nodes in the input and hidden layers are indi-
cated by their constant value of unity
Between this input layer and the single node of the output
layer is a hidden layer with two nodes. The number of nodes in the
hidden layer was chosen equal to the number of classes the input
data needs to distinguish, which are “left fragment charged” and
“left fragment neutral.” The bias nodes of the input and hidden
layers are indicated by their constant value of 1. Although charge
prediction could be treated as classification, we chose to treat it as
multivariable nonlinear regression. Sigmoidal transformations are
used to generate an output between 0 and 1. Back-propagation of
the difference between this output and the label of exemplars was
used to optimize the weights of the neural network with a learning
rate of 0.25 and momentum coefficient of 0.95.
2.3 Training Results The root-mean-square deviation (RMSD) between output values
and exemplar labels was selected as a measure of training conver-
gence. The lower curve in Fig. 6 shows the convergence achieved
by 500 epochs of training. These results reveal that convergence is
rapid, as would be expected for a neural network with only two
nodes in a single hidden layer. The “elbow” in the RMSD for the
training set occurs after about 25 epochs, and the RMSD decreases
very slowly after about 200 epochs.
The upper curve in Fig. 6 shows the RMSD of output values
relative to the label on exemplars of the test set, which was calcu-
lated after every epoch of training. The RMSD relative to the test
set follows the RMSD of the training set for only about 10
epochs, after which it deviates sharply, remains approximately
constant, but increases slightly as the RMSD of the training set
Fig. 6 Root-mean-square deviation (RMSD) of network output from exemplar

labels in training and testing sets as a function of cycles of refinements on the
training set
Table 1
Accuracy of charged-fragment prediction in a test set derived from 50
lipid metabolites distributed over 15 classes (see Note 3) after 500 cycles
of ANN refinement on a similar training set
Correct overall percentage 99.3 %

Incorrect overall percentage 0.7 %
True positives 2,170 (99.4 %)
True negatives 2,166 (99.2 %)
False positives 13 (0.6 %)
False negatives 17 (0.7 %)
continues to decrease. We interpret this behavior as an indication

that there is a small amount of “overfitting” with 500 epochs of
training. Details of the performance of the model on the test set,
after 500 epochs of training, are summarized in Table 1.
2.4 Future Research Results, shown in Table 1, of the machine-learning experiment

with data on 94 lipid metabolites are encouraging; nevertheless,
work is in progress to improve the ANN model for prediction of
charged fragments. This work includes the following: (1) additional
experimental CID spectra of non-lipid metabolites, (2) inclusion
of secondary fragmentation so that data from instruments other
than ion-trap spectrometers can be modeled, (3) larger training
sets with inclusion of cross validation, and (4) more combinations

of training and test sets with the use of random numbers to ensure
uniform representation of all types of metabolites.
3 Notes
1. The 18:0/18:0 notation refers to the number of carbon atoms

and number of double bonds that are in each hydrophobic tail
attached to the lipid’s head group. The unlabeled atoms (i.e.,
numbered only) are understood to be carbon. Hydrogen
atoms needed to saturate chemical bonding are not shown.
More information about the LIPID MAPS database and its
nomenclature can be found at http://www.lipidmaps.org/
data/classification/lipid_cns.html.
2. Details regarding linear ion-trap mass spectrometers and CID
energy calculations can be found in the supplemental material
of Kangas et al. [4].
3. Training and testing exemplar sets were based on experimental
data and molecular structures for the following lipids: phospha-
tidylcholine (14:0/14:0, 14:0/16:0, 16:0/16:0, 16:1/16:1,
17:0/17:0, 18:0/18:0, 18:1/18:1, 18:2/18:2, 18:3/18:3,
20:1/20:1, 20:4/20:4, 23:0/23:0), phosphatidylcholine ether
(13:0/13:0, 18(P)/18:1, 18:1/18:1, 18(P)/20:4), lysophos-
phatidylcholine (14:0, 15:0, 16:0, 17:0, 17:1, 18:0, 18:1),
phosphatidylethanolamine (12:0/12:0, 15:0/15:0, 16:1/16:1,
16:0/18:1, 17:1/17:1, 18:0/18:0, 18:0/18:1), lysophosphati-
dylethanolamine (14:0, 16:0, 18:0, 18:1), phosphatidylserine
(12:0/12:0, 14:0/14:0, 16:0/18:2, 17:0/17:0, 18:0/18:0,
18:0/18:1, 18:0/18:2), lysophosphatidylserine (16:0, 18:0,
18:1), sphingomyelin (d18:1/12:0, d18:1/16:0, d18:1/17:0,
d18:1/18:0, d18:1/24:1), ceramide (d18:1/12:0, d18:1/17:0,
d18:1/18:0, d18:1/20:0, d18:1/22:0, d18:1/24:0), ceramide
1-phosphate (d18:1/8:0, d18:1/12:0, d18:1/16:0, d18:1/18:1,
d18:1/24:0), hexaceramide (d18:1/8:0, d18:1/12:0, d18:1/
16:0, d18:1/24:0, d18:1/24:1), dihexaceramide (d18:1/8:0,
d18:1/12:0), cholesterol, cholesterol ester (18:2, 18:3,20:4.20:5),
diglycerol (17:0/17:0, 20:0/20:0, 20:2/20:2, 20:4/20:4,
20:5/20:5), and triglycerol (14:0/16:1/14:0, 15:0/18:1/15:0,
16:0/18:0/16:0, 16:0/18:1/22:6, 16:1/17:1/17:2, 16:1/
18:1/18:2, 16:1/18:1/20:4, 17:0/17:1/17:0, 18:1/18:1/
18:2, 20:0/20:1/20:0, 20:2/18:3/20:2, 20:4/18:2/20:4,
20:5/22:6/20:5).
4. The complete encoded vector has the form <Meta data>
<Block 1><Block 2>…<Block 8>. The <Meta Data> block,
which includes information about the lipid metabolite under
investigation, is not part of the input attributes but is referred

to in the generation of the output label. MetISIS uses an
undirected graph derived from the chemical structure to rep-
resent a metabolite. When a bond is broken in the CID simula-
tion, the two atoms that were involved in the bond become
root nodes of a disconnected graph consisting of two sub-
graphs. One subgraph is denoted the left fragment and the
other one is denoted the right fragment. The CHONSP count
for each atom within eight bonds of the root atoms in the frag-
ments is easily obtained by a breadth-first search [10].
Each exemplar used in machine learning is associated with
a particular bond cleavage in a specified metabolite. CHONSP
counts for both fragments generated in the cleavage are
retained in the same input vector. Since each peak in the exper-
imental CID spectrum contributes two exemplars to the
charge-prediction exemplar set, a charged and a neutral frag-
ment, a second input vector is generated for each bond cleav-
age by interchanging the left and right components. The
output label of each exemplar identifies the charge (1 or 0) of
the left component. The equal number of charged and neutral
exemplars in the training set prevents the encoding scheme
from introducing bias in charge prediction.
In addition to this basic encoding, MetISIS performs two
more processing steps on the encoded vectors. The first is to
compact the vectors. This is done by examining a large set of
molecules, encoding them as described above and then remov-
ing CHONSP components that are always zero. For example,
if there is no triple-bonded P in any of the molecules at a given
distance N from the root atoms, then the triple-bonded P will
be removed from the corresponding <BlockNP> components.
This mode of compression adapts input vectors to a particular
set of metabolites. For lipid metabolites in Note 3, the com-
pressed input vectors have 1,270 components. The second
step is normalization of the data. After compression, the
remaining CHONSP components are normalized to values
between 0 and 1.
Acknowledgments
Results reported in this chapter were part of masters- and doctoral-

degree programs in computer science at Washington State
University Tri-Cities for B. T. Schrom and L. J. Kangas, respec-
tively. J. H. Miller received partial support from the Low Dose
Radiation Research Program of the US Department of Energy.
References
1. Eng JK, McCormack AL, Yates JR (1994) An database for lipid identification. Nat Methods
approach to correlate tandem mass-spectral 10:755–758
data of peptides with amino-acid-sequences in 6. LIPID Metabolites and Pathways Strategy
a protein database. J Am Soc Mass Spectrom (LIPID MAPS) program. www.lipidmaps.org
5:976–989 7. Goldberg DE (1989) Genetic algorithms in
2. Perkins DN, Pappin DJ, Creasy DM et al search, optimization and machine learning.
(1999) Probability-based protein identifica- Addison-Wesley, Reading, MA
tion by searching sequence databases using 8. Faulon JL, Visco D, Roe D (2005) Enumerating
mass spectrometry data. Electrophoresis 20: molecules. In: Lipkowitz KB et al (eds) Reviews
3551–3567 in computational chemistry, vol 21. Wiley,
3. Scheubert K, Hufsky F, Bocker S (2013) Hoboken, NJ, pp 209–286
Computational mass spectrometry for small 9. Schietgat L, Ramon J, Bruynooghe M et al
molecules. J Cheminform. doi:10.1186/1758- (2008) An efficiently computable graph-based
2946-5-12 metric for the classification of small molecules.
4. Kangas LJ, Metz TO, Isaac G et al (2012) In In: Proceedings of the 11th international con-
silico identification software (ISIS): a machine ference on discovery science (LNAI 5525),
learning approach to tandem mass spectral pp 197–209
identification of lipids. Bioinformatics 28: 10. Cormen TH, Leiserson CE, Rivest RL et al
1705–1713 (2009) Elementary graph algorithms. In:
5. Kind T, Liu KH, Lee do Y et al (2013) Introduction to algorithms, 3rd edn. MIT
LipidBlast in silico tandem mass spectrometry Press, Cambridge, pp 594–602
Chapter 7
Prediction of Bioactive Peptides Using Artificial

Neural Networks
David Andreu and Marc Torrent
Abstract
Peptides are molecules of varying complexity, with different functions in the organism and with remarkable
therapeutic interest. Predicting peptide activity by computational means can help us to understand their
mechanism of action and deliver powerful drug-screening methodologies. In this chapter, we describe how
to apply artificial neural networks to predict antimicrobial peptide activity.
Key words Artificial neural network, Peptide, Antimicrobial, Screening, Drug
1 Introduction
1.1 Peptides Peptides are natural or synthetic polymers composed of amino

and Their Use acids linked by an amide bond. The different amino acids have
in Medicinal different physicochemical characteristics. In nature, most proteins
Chemistry and peptides are built from 20 natural amino acids, which are
depicted in Fig. 1a.
The peptide chain has a reduced conformational flexibility due
to a significant double-bond character between the carbonyl car-
bon and the nitrogen (Fig. 1b). The peptide chain is mainly planar
and conformational changes are restricted, with flexibility mainly
limited to rotations around the α-carbon atoms [1]. In addition,
physicochemical constraints, such as side-chain volume and charge,
contribute to peptide structure. Therefore, peptides can be either
unfolded or structured in α-helix, β-strands with loops and coils
(Fig. 1c), or a combination of them. Globally, the amino acid com-
position, together with the three-dimensional structure, is what
determines peptide action.
Electronic Supplementary Material: The online version of this chapter (doi: 10.1007/978-1-4939-2239-0_7)
contains supplementary material, which is available to authorized users
101
102 David Andreu and Marc Torrent
a
Positively charged Negatively charged
O O O O O
NH NH NH NH NH
CHARGED
OH
N O O
NH
NH OH
HN H2N
NH 2
Arginine (Arg) Histidine (His) Lysine (Lys) Aspartic (Asp) Glutamic (Glu)
Polar uncharged Special amino acids
O O O O O O O
NH NH NH NH NH NH NH
POLAR
HO
OH CH3 NH 2 SH
O O
NH 2
Serine (Ser) Threonine (Thr) Asparagine (Asn) Glutamine (Gln) Glycine (Gly) Serine (Ser) Proline (Pro)
Hydrophobic aliphatic Hydrophobic aromatic

O O O O O O O
HYDROPHOBIC
NH NH NH NH NH NH NH
H3C H3C H3C
CH3
H3C S NH
CH3
OH
Alanine (Ala) Valine (Val) Isoleucine (Ile) Methionine (Met) Phenilalanine (Phe) Tyrosine (Tyr) Tryptophan (Trp)
b
R1 O R3 O
H H H H
N N
N N
H R2 H H
O H O R4
c
O R2 O R4
H H
H H
N N
N N
H H H H
R1
NH
R1 O R3 O
H
O
R3
O R3 O R1
H H H H
HN
O
N N
H
N N
H H H
H
R4 O R2 O
Fig. 1 Peptide composition and structure. Peptides are built from amino acids with different chemical properties
(a) linked by amide bonds. The peptide chain is mainly planar, with flexibility mostly limited to rotations
around the α-carbon atoms (b). Despite these constraints, peptides can adopt different secondary structures
(c) including α-helix and β-strands with loops and coils
Prediction of Bioactive Peptides 103
Table 1
Selected examples of natural peptides
Peptide name Peptide sequence Function

Glutathione γ-Glu-Cys-Gly Antioxidant
Angiotensin II Asp-Arg-Val-Tyr-Ile-His-Pro-Phe Vasoconstrictor
Bradykinin Arg-Pro-Pro-Gly-Phe-Ser-Pro-Phe-Arg Vasodilator
Calcitonin Cys-Gly-Asn-Leu-Ser-Thr-Cys-Met-Leu-Gly-Thr-Tyr-Thr- Calcium metabolism
Gln-Asp-Phe-Asn-Lys-Phe-His-Thr-Phe-Pro-Gln-Thr-Ala-
Ile-Gly-Val-Gly-Ala-Pro
Somatostatin Ala-Gly-Cys-Lys-Asn-Phe-Phe-Trp-Lys-Thr-Phe-Thr-Ser-Cys Growth and cell
division
Melittin Gly-Ile-Gly-Ala-Val-Leu-Lys-Val-Leu-Thr-Thr-Gly-Leu-Pro- Antimicrobial peptide
Ala-Leu-Ile-Ser-Trp-Ile-Lys-Arg-Lys-Arg-Gln-GlnNH2
Enkephalin Tyr-Gly-Gly-Phe-Met/Leu Nociception
There are different sources of natural peptides of varying

complexity with different functions in the organism [2]. For exam-
ple, glutathione is an antioxidant peptide that protects cells
from free radicals [3], and angiotensin is a hormone that acts as a
vasoconstrictor, increasing blood pressure [4]. Some well-known
natural peptides are listed in Table 1.
The number of possible combinations of amino acids into pep-
tides is huge and depends upon the peptide length. For example,
for a peptide with 15 amino acids, we could build 2015 different
molecules, meaning more than 1019 different combinations.
Advanced synthetic strategies now allow us to build hundreds or
even thousands of peptides easily [5–8]. Thanks to these technolo-
gies, we are not restricted to natural peptide templates and can
design new peptides with novel functions that do not occur in nature
but have relevant applications in pharmacology and medicine. Given
the number of possible different combinations of amino acids, the
biomedical applications of peptides are nearly infinite.
Unfortunately, we are far from exploring the effects of millions
and millions of amino acid combinations. Even though modern
screening methods are very versatile and widespread, testing large
libraries of compounds can be time intensive, especially when cellular
or animal models are required [9, 10]. In these situations, mathemati-
cal models are essential to narrow down the experimental work
required. In addition, these models help us to understand peptide
activity and can be used to dissect the molecular properties of active
compounds and discover the receptors and pathways in which they are
involved [11]. For all these reasons, mathematical models are essential
in peptide design and have real relevance in the pharmaceutical indus-
try for the development of new antibiotics or anti-inflammatory drugs.
1.2 Use of Prediction As detailed in Table 1, peptides can have diverse functions, which
Models in the Search are correlated with amino acid composition and structure. It is of
for New Drugs particular interest to develop computational tools that are able to
predict active peptides, based on their physicochemical and struc-
tural characteristics. In this chapter we will focus on antimicrobial
peptide classification, but the methods and techniques described
here are widely applicable. Therefore, the same strategies can be
applied in the prediction of other classes of biologically active
peptides.
We aim to describe in this chapter how one can use artificial
neural networks (ANNs) to predict peptide activity. We will
describe how to build and curate datasets, train an ANN, and vali-
date the results in the software package R.
2 Materials
1, 2. Software package R (see Note 1) and software package

RStudio (see Note 2).
3. Libraries (Interpol, corrplot, caret, RSNNS, RColorBrewer,
SDMTools, devtools, reshape) (see Note 3).
4. ANN script (Supplementary Materials).
5. Peptide dataset AMPlist (Supplementary Materials).
6. Descriptor list (Supplementary Materials).
3 Methods
In this chapter we will describe the basics of peptide classification

using ANNs. In our example, we will develop a tool to classify
peptides as antimicrobial or non-antimicrobial.
The protocol is structured in the following steps:
1. Dataset building
2. Descriptor selection
3. Network training
4. Model validation
5. Network visualization
All these steps are covered in the following points with tips and
hints that can be used to apply the present model (or an extension
of it) to analyze similar problems.
3.1 Build Datasets Though it might appear obvious, it is worth stating that the results
of computational analysis are strongly dependent on the quality of
the databases provided. For a supervised learning strategy such as the
one described in this chapter, we need to provide a list of peptides

(input) and a measure of the biological property of interest for
each peptide (output). In the classification example we are present-
ing here, the outputs are in a yes/no format, denoting that the
peptide has or does not have antimicrobial properties respectively.
When using data from one’s own laboratory, it is possible to fix
a set of criteria that will be applied to all samples, which are then
processed using the same protocols. However, when obtaining
data from the literature, caution is required to ensure consistency.
Previously published data from other laboratories may have been
gathered using many different protocols. For example, different
media for bacteria cultures can affect the assessment of antimicro-
bial properties. Whenever possible, it is wise to use only values
obtained from the same method.
When this is not possible, it may be necessary to include data
from other sources. For antimicrobial peptides and bioactive pep-
tides in general, there are public databases that gather information
about peptides. In general, the classification criterion in these data-
bases is based on bibliography searches, without taking into
account the concerns described above. It is important to keep in
mind that this can affect the performance of the calculations and
that as a result less accurate prediction systems may be obtained.
We describe in Subheading 3.1.1 how to build a positive dataset
from public databases.
It is also important to avoid overrepresentation of particular
peptide families. When a single family of peptides represents a large
part of the dataset, this may bias the results. Always check for pep-
tide similarity and use a filter as described in Subheading 3.1.2 to
remove large groups of similar peptides.
Though positive databases can be difficult to obtain, more
serious problems can arise from the need for a negative data-
base, because it is hard to find a detailed list of negative results.
For example, numerous antimicrobial peptides can be found in
the literature but only a few examples of non-antimicrobial pep-
tides are described. This is a major problem, as it may bias the
training process.
Frequently, large negative databases may not be built because
not enough data is available. Therefore, an alternative method is
required to obtain a suitable negative dataset. This is the case of
antimicrobial peptides. Though public databases contain around
2,000–5,000 antimicrobial peptides (positive database), it is
impossible to build an experimentally validated negative dataset
of similar size for non-antimicrobial peptides. In Subheading
2.1.2 we describe a method to build such a dataset. Though this
approach may be useful in antimicrobial peptides, it may not be
applicable to other bioactive peptides, and particular strategies
may need to be found, depending on the type of study.
Table 2
Databases of antimicrobial peptides
Database Number of entries Data source References

APD(APD2) 2,338 Pubmed [12]
BACTIBASE (BACTIBASE 2) 177 Uniprot [13]
Pubmed
CAMP 4,020 NCBI [14]
Defensins knowledgebase 350 SciFinder Scholar [15]
Uniprot
EMBL/GenBank
PenBase 110 Pubmed [16]
EMBL/GenBank
PhytAMP 271 Uniprot [17]
Pubmed
RAPD 179 Pubmed [18]
3.1.1 Build the Positive The positive dataset is built from a list of peptides which have been
Dataset shown experimentally to have the property of interest. The elements
of the dataset can be retrieved from experiments done in the labora-
tory or from public databases. The dataset consists of a text file built
using either a text engine or a spreadsheet containing the list of
peptides. It should be saved as a non-formatted text file (.txt exten-
sion) and the peptides should be delimited by a carriage return.
For antimicrobial peptides, useful databases are listed in Table 2.
In our working example, we have selected the cationic peptides
from the APD2 database (see AMPlist.txt included in the
Supplementary Material).
3.1.2 Build the Negative This dataset contains a list of peptides validated to not have the
Dataset property of interest. As was the case with the positive dataset, it can
be built based on experiments performed in the laboratory, or by
relying on public databases, but also by extrapolation as described
in the introduction to this section. The dataset consists of a text file
as described in Subheading 2.1. Like the positive dataset, it should
be saved as a non-formatted text file (.txt extension) and the pep-
tides should be delimited by a carriage return.
To build our example negative dataset, we have used the
Uniprot server (http://uniprot.org) to search for non-antimicrobial,
nontoxic, and non-antibiotic peptides, with restricted length (from
5 to 45 amino acids). Only Uniprot-refined entries were selected.
A list of >3,000 peptides was finally included in the negative
dataset. As these peptides may have high sequence similarity, we
used CD-Hit (http://weizhong-lab.ucsd.edu/cdhit_suite/, [19])
to reduce sequence redundancy (90 %). Our final negative dataset

contained 2,234 peptides. We have also manually curated the data-
base to delete protamines and histones that are erroneously not
categorized by Uniprot as antimicrobials (see AMPlist.txt included
in the Supplementary Material).
3.1.3 Merge Datasets Merging the datasets is an optional but convenient step. To merge
and Add the Class Vector the datasets, we have merely copied the text contained in the posi-
tive and negative datasets (in that order) into a new text file, called
AMPlist (see AMPlist.txt included in the Supplementary Material).
To train the classification system, peptides in the list must be
labeled as active or inactive. To do that, active peptides were labeled
1 and inactive peptides labeled 0. While a list of 0 s and 1 s can be
built using a spreadsheet and then loaded into the training soft-
ware, we have for the sake of convenience included instead a code
line to build the class vector (see below).
3.1.4 Script Explanation The script provided uses the R function read.table() to load the pep-
tide list file. In the example, the peptide list called AMPlist.txt has a
list of 2,048 antimicrobial peptides and 2,234 non-antimicrobial
peptides, built as described above. To load a new list instead of the
example, just change the name in the script (line 13, bold text):
peptide_list<- read.table("AMPlist.txt", quote="\"")
The class vector, containing an appropriate number of 0 s and
1 s, is created in line 19 of the script:
y<- c(rep(1,2048),rep(0,2234))
The first parenthesis contains the number of active peptides
and the second parenthesis the number of inactive peptides.
Numbers can be changed but must be consistent with the list sup-
plied as peptide_list.
3.2 Select Molecular descriptors are numerical values that characterize prop-
Descriptors erties of the molecule of interest [20]. For example, a simple
molecular descriptor could be the molecular weight or the charge
of the peptide. There are many types of molecular descriptors, but
they can be classified as either (1) structure-based, which depend
on the structural properties and include topological or geometrical
descriptors, or (2) property-based, which are experimentally or
theoretically determined values and include properties such as
hydrophobicity or accessible surface area.
Structure-based descriptors can refer to different levels of
organization. For example, in peptides we can refer to (1) the
amino acid sequence, (2) the secondary structure of particular
segments in the peptide, and (3) the complete tertiary structure of
the peptide.
Most commonly, a three-dimensional structure is not available,

so no experimentally validated values can be obtained for structure-
based descriptors. At this point, structure modeling can be used,
but this is time consuming and generally not reliable. Therefore,
in this chapter we will only consider sequential descriptors and
simple predictions of secondary structure elements (e.g., the pro-
pensity of the peptide to adopt an α-helix structure).
3.2.1 Descriptor One of the most extensive lists of descriptors for amino acids is
Calculation encoded in the AAindex database [21, 22]. The AAindex database
contains a selected list of 533 descriptors, based on amino acid
physicochemical properties, substitution matrices, and statistical
protein contact potentials.
The method of calculating properties based on the AAindex
values is also flexible. The simplest strategy is to compute the aver-
age for all amino acids present in the peptide sequence. This
approach is computationally inexpensive but may miss important
details like the order in which the amino acids appear in the struc-
ture. For example, the peptide ASLP will have the same value as
the peptide PALS, though the peptides may have completely dif-
ferent biological properties.
Another approach is to use autocorrelation, which takes into
account the distribution of these properties along the sequence of
peptides [23, 24]. There are several strategies to compute autocor-
relation; one example is the Moreau-Broto autocorrelation:
N -d
AC MB = åP P i i +d
i =1
where the property of interest is named P and the autocorrelation
value is computed as the summation of products along the sequence
for amino acids separated by a length d. For the example described
before, assuming a distance d of 1 amino acid and the arbitrary
values A = 1, S = 2, L = 3, and P = 4, we would have an ACMB = 1 × 2
+ 2 × 3 + 3 × 4 = 20 for peptide ASLP and AC = 4 × 1 + 1 × 3 + 3 × 2 = 13
for peptide PALS.
In the script provided in this chapter, we will use the average
strategy, but autocorrelation can be also implemented in the script.
To learn more about other methods to calculate autocorrelation,
see refs. 23, 24.
Though it is possible to calculate all descriptors in the script
provided, several limitations must be taken into account. First, the
number of descriptors should be small compared with the number
of examples provided for the training process to be accurate. For
example, it is not suitable to use 400 descriptors if the peptide
dataset has only 200 peptides. Second, using a disproportionate
number of descriptors could easily lead to model over-fitting. Third,
many descriptors are highly correlated with others, generating
redundancy in the data. For example, the AAindex list of descriptors

contains several different measures of amino acid hydrophobicity.
Fourth, the more complex the model, the more difficult it is to
analyze the results. When using a reduced number of parameters,
it is easy to understand the contribution of each of them to the
model and to understand its biological relevance. When large
numbers of parameters are used, their contribution tends to bal-
ance and details are easily lost (see Note 4).
3.2.2 Script Explanation In the script provided, only a small subset of parameters [13] is used
for prediction. To load a different subset of parameters, substitute
the numbers given in lines 22 and 27 (highlighted in bold) by the
corresponding numbers in the descriptor list (see Supplementary
Material):
vector<- as.numeric(c("9", "37", "38", "39", "40", "69", "71",
"96", "146", "151", "319", "321", "381", "401"))
rownames(ff)<- c("AA9", "AA37", "AA38", "AA39", "AA40",
"AA69", "AA71", "AA96", "AA146", "AA151", "AA319",
"AA321", "AA381", "AA401")
In lines 30–39, we implement a code for deleting highly cor-
related parameters. Therefore, those parameters that display a cor-
relation index of 0.90 with another parameter are disregarded for
the network training (see Note 5).
The threshold can be changed in the script, in line 33:
highlyCor<- findCorrelation(cor2, 0.90)
3.3 Develop Artificial neural networks are computational models inspired in the
the Artificial Neural brain neuronal system [25, 26]. The ANNs consist of inputs (rep-
Network Model resenting the signal received by the neuron) that are multiplied by
weights (representing the strength of the signal) and then mathe-
matically transformed to produce the outputs (the response to the
input signal). This is represented schematically in Fig. 2a.
In summary, ANNs can be represented as weighted directed
graphs with neurons represented by nodes and connections rep-
resented by directed edges. Following this representation, ANNs
can be classified as (1) feed-forward networks or as (2) recurrent
(or feedback) networks. The difference between them is that the
first does not contain any loops while the later does contain
loops (Fig. 2b). In this chapter we will focus only on feed-for-
ward networks and, concretely, on multilayer perceptron net-
works, a particular ANN in which neurons are organized in layers
and are connected with unidirectional edges (Fig. 2c).
In this chapter, we aim to build ANNs that learn from given
examples (known active and inactive peptides) to be able to predict
new inputs (unknown peptides). This scheme is known as supervised
learning because we provide a set of examples to help the network to
a
Input 1
w1
Input 2 w2
w3 T OUTPUT
Input 3
w4
Input 4
b
Input 1
T OUTPUT
Input 2
Input 1
T OUTPUT
Input 2
c
Input 1
Input 2
T
Input 3
Input 4
T OUTPUT
Input 5
Input 6
T
Input 7
Input 8
Fig. 2 Typical representation of artificial neural networks. Artificial neural net-

works consist of weighted inputs that are mathematically transformed to pro-
duce the outputs (a). These networks can be represented as weighted directed
graphs and can be classified as feed-forward networks and recurrent (or feed-
back) networks (b). Concretely, in multilayer perceptron networks, neurons are
organized in layers connected by unidirectional edges (c)
learn from them, in a process called network training. The network

learns how to reproduce the outputs by modifying the weights that
multiply the inputs.
To successfully train a network, we have to provide an environ-
ment in which the network will operate (called the learning para-
digm) and the rules that the network will use to update the weights
(called the learning algorithm).
In our case, the learning paradigm is defined by peptide
sequences and measurable properties (physicochemical parameters,
biological activity, etc.). The learning rule will be an error-
correction rule. In this rule, we randomly initialize the weights and
feed the network with a training vector (a list of peptides with their
properties and the classification index, 1 for active and 0 for inac-
tive). The network will then update the weights in order to mini-
mize the difference between the calculated and the given output
(cost function). To do that, we use the back-propagation method
that incorporates gradient descent to minimize the squared-error
cost function.
Several packages are available to simulate neural network mod-
els in R, including neuralnet [27], RSNNS [28], and CARET [29]
among others. Among the advantages of the CARET package are
that it can run most network models using the same syntax and can
automatically apply different pretreatment methods (e.g., principal
component analysis) and validation protocols (e.g., k-fold cross
validation). Therefore, we will be using the CARET package in our
example.
3.3.1 Selection In the script provided, we have divided the entire dataset (AMPlist)
of the Training and Test into a training set that contains 75 % of the observations and a test-
Datasets ing set that contains 25 % of the total observations. This division is
variable and can be changed in line 46 (bold number) (see Note 6):
smp_size<- floor(0.75 * nrow(M))
3.3.2 Training The training step consists of feeding the model with input data and
and Validation allowing it to optimize the weights in order to find the values that
best fit the data. Different training strategies have been described
in Subheading 2.3, and therefore, in this section, we will focus on
explaining how validation is performed and why it is important.
Validation is an essential step in assessing the strength of the
method. If the model were trained with the entire set of input data,
we could expect that it would find a set of weights that fit the data
with high accuracy. However, when this model is tested with a new
set of examples, previously unseen by the model, we may find a
poor prediction accuracy. This behavior is called over-fitting, which
is a computational artifact that does not generate a true fitting rule
but an overcompensated model that only fits the example data.
A validation step is thus required to evaluate when over-fitting

is happening. The simplest form of validation consists of a k-fold
cross validation. In this method we divide the dataset into k differ-
ent equal-sized sets. In k different training rounds, one set is used
as the testing set while the remaining sets are used for training the
model. The average error and standard deviation of each round
provides an estimation of how robust is the method.
A particular k-fold cross-validation strategy is called leave-one-
out cross-validation strategy. This is achieved when k is equal to the
number of observations minus one; then, only one prediction is
made in each round (see Note 7).
As detailed before, in our example we use a multilayer percep-
tron (mlp) to fit our data using the CARET package in R. This
package allows multiple options (see the CARET documentation
for further details [29]).
Interestingly, we can specify the validation method through
the trControl instruction. In our script, the validation strategy is
fivefold cross validation, as detailed in line 56:
trC=trainControl(method="cv", number=5)
Multilayer perceptron networks can have different neuron
layers (called hidden layers) with a variable number of neu-
rons. We can specify this using the tuneGrid instruction pro-
vided in the caret package. In our script we have used only one
hidden layer with a variable number of neurons (line 57, from
0 to 5):
my.grid<- expand.grid(.size=c(0,1,2,3,4,5))
Training is specified in the CARET package by using the train
function. For the train function to be correctly defined, we must
specify the training data, the classification data, the method used
for training, the preProcess strategy, the validation strategy, and
the variations in the tuneGrid parameter. The training instruction
is detailed in line 58 of the script provided:
network <- train(train[,1:9], train[,10], method="mlp",
preProcess=NULL, trControl=trC, tuneGrid=my.grid)
We have reproduced here typical output for a multilayer per-
ceptron trained as described above:
3,211 samples
9 predictors
No pre-processing
Resampling: cross validation (fivefold)
Summary of sample sizes: 2,569, 2,569, 2,569, 2,569, 2,569
Resampling results across tuning parameters:
Size RMSE Rsquared RMSE SD Rsquared SD

0 0.41 0.366 0.0235 0.0477
1 0.406 0.376 0.0102 0.0409
…
4 0.358 0.515 0.0185 0.0386
5 0.346 0.539 0.00706 0.0187
RMSE was used to select the optimal model using the smallest
value. The final value used for the model was size = 5.
The results output summarize the method, meaning that we
have used 9 predictors over a database of 3,211 examples, with no
pre-processing and with a fivefold cross-validation strategy (each
step comprising 2,569 examples used to train the method).
The method has tested the different number of neurons in the
hidden layer, from 0 to 5. To determine which model was best, the
method identifies the model with the lowest root mean square
error (RMSE), which is the model with 5 neurons. The RMSE is
the square root of the variance of the residuals, indicating the abso-
lute fit of the model to the data. For classification purposes, our
model outputs have values between 0 and 1 while the true values
are binarized (either 0 or 1). RMSE measures the error distance
between predicted and real data. As we need a threshold to map
predicted to real data, RMSE does not account for prediction
accuracy but is a good indicator of how well the model fits the data.
As a rule of thumb, the lower the better; it should not be higher
than 0.5.
3.3.3 Testing We use the isolated section of the dataset (in our case the 25 %) to
test the prediction power of the method. In the testing step, the
method is applied to the testing dataset and the predictions are
compared with the real values. In the best-case scenario, the pre-
dicted values will be identical to the real values. We define true posi-
tive values (TP) as those positive observations that are predicted as
positive (in our case, antimicrobial peptides correctly predicted
as antimicrobial) and true negatives (TN) as negative observations
that are correctly predicted as negative (non-active peptides pre-
dicted as non-antimicrobial). We can also have false positives (FP,
non-active peptides predicted as antimicrobials) and false negatives
(FN, active peptides predicted as non-antimicrobials).
Therefore, we can build a matrix, as depicted in Table 3, known
as the confusion matrix. Good prediction methods will have a large
number of observations as true positives or true negatives and low
numbers of false negatives and false positives.
Table 3
Confusion matrix
Prediction
0 1
Real 0 TN FP
1 FN TP
Good indicators of successful methods include a good sensitivity

(the proportion of positives that are true) and specificity (the ability
of the method to identify negative results that are truly negative).
The mathematical expression of these indicators is
TP
Sensitivity =
TP + FN
TN
Specificity =
TN + FP
Accuracy can be defined as the ability to predict true outcomes
(positives and negatives) and can be calculated as
TP + TN
Accuracy =
TP + FP + TN + FN
In binary classifications (e.g., our case to classify peptides in anti-

microbial and non-antimicrobial), we can use the Matthews cor-
relation coefficient as a measure of the prediction quality. This
coefficient is, briefly, a measure of the correlation between the
observations and predictions. It can be calculated using the follow-
ing formula:
TP ´ TN - FP ´ FN
MCC =

( TP + FP ) ´ ( TP + FN ) ´ ( TN + FP ) ´ ( TN + FN )
Finally, when possible, an experimental validation of the method
can also be performed as the ultimate test to validate the strategy
presented. This is accomplished by generating a new list of com-
pounds, not previously seen by the algorithm and analyzed by the
method. Then a subset of predictions containing both positive and
negative examples is experimental validated.
In the script provided, the confusion matrix, specificity, selec-
tivity, and accuracy are calculated in lines 63 and 64 of the script
(see Note 8):
confusion.matrix(test3, pred, threshold=0.5)
accuracy(test3, pred, threshold=0.5)
The results obtained in our network example give around 85 %

sensitivity, 85 % specificity, and 85 % accuracy. This is a good result,
as the network detected 85 % of all positive examples in the testing
dataset and 85 % of the negative results were true negatives.
Despite this, there is still room for improvement; this might
be accomplished by a variety of different means. Possible strategies
to obtain a better classification are: (1) include more parameters,
particularly those covering other physicochemical characteristics,
(2) include more neurons in the hidden layer, and (3) include
another hidden layer.
It is difficult to suggest a priori which is the best strategy to
follow but it is important to remember that all the strategies will
add one or more degrees of complexity, therefore complicating the
interpretation of the results. A compromise must always be consid-
ered between how important it is to get good predictions and how
important it is to understand how the model works.
3.4 Network To help understand the structure of the neural network and how it
Visualization operates, it is useful to represent it as shown in Fig. 2. Also, it is
important to specify the weights, that is, the contribution of each
node to the next layer of the network. The higher the contribution
of the node, the more important it is to the prediction.
In the script provided (line 69), we have used a function called
plot.nnet to visualize our network (please see http://beckmw.
wordpress.com/ for a complete description of this function):
Repnet<- mlp(train[,1:9], train,10, size=4)
plot.nnet(repnet)
First, we recompute the neural network with a number of hid-
den neurons (size) specified by the output results as detailed in
Subheading 3.3.2 and then use the plot.nnet function to plot the
results (Fig. 3).
Input_1 I1
Input_2 I2
Input_3 I3
H1
Input_4 I4
H2
Input_5 I5 O1 Output_1
H3
Input_6 I6
H4
Input_7 I7
Input_8 I8
Input_9 I9
Fig. 3 Example of a neural network obtained using the example model proposed in this chapter
Weights are depicted in the edges that connect the nodes. The
absolute value of the weight is symbolized by the width of the line
(the wider the line, the higher the contribution). The line color
symbolizes whether the contribution is positive (green) or negative
(blue). In this example, descriptors 6 and 7 are contributing sig-
nificantly to the model while descriptor 2 is contributing less.
The plot.nnet function does not depict the bias values (con-
stant values added to each hidden and output layer) for the mlp
neural network; however, this has no impact in the interpretation
of our results. The plot.nnet function only depicts the bias values
for the nnet function.
4 Notes
1. The R package can be downloaded from http://www.r-project.

org/.
2. The RStudio package can be downloaded from http://www.
rstudio.com/.
3. To install the required libraries in R, use the install.pack-
ages(“”) function.
4. Our suggestion is always to select a number of parameters that
are thought to be relevant and then, if the results are not satis-
factory, try other parameters or extend the list by a small
amount.
5. A threshold between 0.9 and 0.95 can be used normally. If the
threshold is much higher than 0.95, only a limited degree of
filtering will be achieved; on the other hand, if it is less than
0.9, there is a risk that information that may be important for
the network can be lost.
6. Care should be taken not to use too small a testing set, since
that will lead to nonrepresentative results. At the other extreme,
a very large testing set will leave too few observations for use
during training and the model will fit poorly. In general, this
parameter can be adjusted between 0.6 and 0.9 depending on
the total number of total observations though values between
0.75 and 0.85 are recommended.
7. When the database is large, it is time consuming to perform a
leave-one-out cross-validation approach because many training
cycles will be required for large datasets. In that case, a more
suitable strategy is to validate the method using a k-fold cross
validation, usually with k values ranging from 4 to 10.
8. Because the activation function in the neural network is the
logistic function, the output values are in the range [0,1]. The
threshold value will binarize the results. Therefore, if the
threshold is 0.5, values below 0.5 will be assigned the output
value 0 and above 0.5 the output value 1.
Acknowledgements
This work was supported by the European Union [grant number

HEALTH-F3-2008-223414 to D.A. and FP7-PEOPLE-2012-
IEF-330352 to M.T.], the Ministerio de Economía y Competitividad
[grant number SAF2011-24899 to D.A.], and the Generalitat de
Catalunya [grant number SGR2009-494 to D.A.]. M.T. would also
like to acknowledge support from the Ramón y Cajal Program
(Spain).
References
1. Nelson DL, Cox MM (2005) Lehninger prin- 12. Wang G, Li X, Wang Z (2009) APD2: the
ciples of biochemistry. 4th edn. p 1100 updated antimicrobial peptide database and its
2. Liu F, Baggerman G, Schoofs L et al (2008) application in peptide design. Nucleic Acids
The construction of a bioactive peptide data- Res 37:D933–D937
base in Metazoa. J Proteome Res 13. Hammami R, Zouhir A, Le Lay C et al (2010)
7:4119–4131 BACTIBASE second release: a database and
3. Zhang H, Forman HJ (2012) Glutathione syn- tool platform for bacteriocin characterization.
thesis and its role in redox signaling. Semin BMC Microbiol 10:22
Cell Dev Biol 23:722–728 14. Waghu FH, Gopi L, Barai RS et al (2014)
4. Patel BM, Mehta AA (2012) Aldosterone and CAMP: collection of sequences and structures
angiotensin: role in diabetes and cardiovascular of antimicrobial peptides. Nucleic Acids Res
diseases. Eur J Pharmacol 697:1–12 42:D1154–D1158
5. Amblard M, Fehrentz J-A, Martinez J et al 15. Seebah S, Suresh A, Zhuo S et al (2007)
(2005) Fundamentals of modern peptide syn- Defensins knowledgebase: a manually curated
thesis. Methods Mol Biol 298:3–24 database and information source focused on
6. Coin I, Beyermann M, Bienert M (2007) the defensins family of antimicrobial peptides.
Solid-phase peptide synthesis: from standard Nucleic Acids Res 35:D265–D268
procedures to the synthesis of difficult 16. Gueguen Y, Garnier J, Robert L et al (2006)
sequences. Nat Protoc 2:3247–3256 PenBase, the shrimp antimicrobial peptide
7. Hilpert K, Winkler DFH, Hancock REW penaeidin database: sequence-based classifica-
(2007) Peptide arrays on cellulose support: tion and recommended nomenclature. Dev
SPOT synthesis, a time and cost efficient Comp Immunol 30:283–288
method for synthesis of large numbers of pep- 17. Hammami R, Ben HJ, Vergoten G et al (2009)
tides in a parallel and addressable fashion. Nat PhytAMP: a database dedicated to antimicro-
Protoc 2:1333–1349 bial plant peptides. Nucleic Acids Res 37:
8. Winkler DFH, Campbell WD (2008) The spot D963–D968
technique: synthesis and screening of peptide 18. Li Y, Chen Z (2008) RAPD: a database of
macroarrays on cellulose membranes. Methods recombinantly-produced antimicrobial pep-
Mol Biol 494:47–70 tides. FEMS Microbiol Lett 289:126–129
9. Reddy AS, Pati SP, Kumar PP et al (2007) Virtual 19. Huang Y, Niu B, Gao Y et al (2010) CD-HIT
screening in drug discovery—a computational Suite: a web server for clustering and compar-
perspective. Curr Protein Pept Sci 8:329–351 ing biological sequences. Bioinformatics 26:
10. Boix E, Nogues VM, Torrent M (2012) 680–682
Discovering new in silico tools for antimicro- 20. Jenssen H (2011) Descriptors for antimicro-
bial peptide prediction. Curr Drug Targets bial peptides. Expert Opin Drug Discov 6:
13:1148–1157 171–184
11. Torrent M, Andreu D, Nogués VM et al (2011) 21. Kawashima S, Pokarowski P, Pokarowska M
Connecting peptide physicochemical and anti- et al (2008) AAindex: amino acid index data-
microbial properties by a rational prediction base, progress report 2008. Nucleic Acids Res
model. PLoS One 6:e16968 36:D202–D205
22. Kawashima S, Ogata H, Kanehisa M (1999) 25. Krogh A (2008) What are artificial neural net-
AAindex: amino acid index database. Nucleic works? Nat Biotechnol 26:195–197
Acids Res 27:368–369 26. Heaton J. (2012) Introduction to the Math of
23. Li ZR, Lin HH, Han LY et al (2006) Neural Networks. Heaton Research Inc. 119
PROFEAT: a web server for computing struc- pages.
tural and physicochemical features of proteins 27. Günther F, Fritsch S (2010) neuralnet: training
and peptides from amino acid sequence. of neural networks. R J 2:30–38
Nucleic Acids Res 34:W32–W37 28. Bergmeir C, Benitez JM (2012) Neural Networks
24. Rao HB, Zhu F, Yang GB et al (2011) Update in R using the Stuttgart Neural Network
of PROFEAT: a web server for computing Simulator: RSNNS. J Stat Software 46:1–26
structural and physicochemical features of pro- 29. Kuhn M (2008) Building predictive models in
teins and peptides from amino acid sequence. R using the caret package. J Stat Software 28:
Nucleic Acids Res 39:W385–W390 1–26
Chapter 8
AutoWeka: Toward an Automated Data Mining

Software for QSAR and QSPR Studies
Chanin Nantasenamat, Apilak Worachartcheewan, Saksiri Jamsak,
Likit Preeyanon, Watshara Shoombuatong, Saw Simeon, Prasit Mandi,
Chartchalerm Isarankura-Na-Ayudhya, and Virapong Prachayasittikul
Abstract
In biology and chemistry, a key goal is to discover novel compounds affording potent biological activity or
chemical properties. This could be achieved through a chemical intuition-driven trial-and-error process or
via data-driven predictive modeling. The latter is based on the concept of quantitative structure-activity/
property relationship (QSAR/QSPR) when applied in modeling the biological activity and chemical prop-
erties, respectively, of compounds. Data mining is a powerful technology underlying QSAR/QSPR as it
harnesses knowledge from large volumes of high-dimensional data via multivariate analysis. Although
extremely useful, the technicalities of data mining may overwhelm potential users, especially those in the
life sciences. Herein, we aim to lower the barriers to access and utilization of data mining software for
QSAR/QSPR studies. AutoWeka is an automated data mining software tool that is powered by the widely
used machine learning package Weka. The software provides a user-friendly graphical interface along with
an automated parameter search capability. It employs two robust and popular machine learning methods:
artificial neural networks and support vector machines. This chapter describes the practical usage of
AutoWeka and relevant tools in the development of predictive QSAR/QSPR models. Availability: The
software is freely available at http://www.mt.mahidol.ac.th/autoweka.
Key words Quantitative structure-activity relationship, Quantitative structure-property relationship,

QSAR, QSPR, Data mining
1 Introduction
Interactions of drugs with their target proteins are governed by a

multitude of molecular forces including electrostatic, hydropho-
bic, polar, and steric. They can be rationalized through the use of
the quantitative structure-activity/property relationship (QSAR/
QSPR) paradigm, which has been one of the foremost tools in the
arsenal of recent drug discovery efforts and has been almost 100
years in the making. The prelude to the formulation of QSAR/
QSPR was the preliminary and independent efforts of Brodin [1]
119
120 Chanin Nantasenamat et al.
and Cros [2] in the mid-1850s; they established that there exists an
association between chemical constitution and physiological prop-
erties. This was at a time prior to the reporting of the aromatic ring
structure of benzene in 1865 by Kekulé [3]. In 1868, Brown and
Fraser observed changes to physiological action upon methylation
of the basic nitrogen atom in alkaloids. In 1869, Richardson [4]
showed that aliphatic alcohols of different molecular weight also
displayed varied narcotic effects. Similarly, a few decades later in
1893, Richet [5] showed that the water solubility of polar chemi-
cals (such as alcohols, ethers, and ketones) was inversely related to
their toxicities, whereby a decrease in the solubility of a polar
chemical results in an increase in its toxicity. Toward the 1900s,
Meyer and Overton [6, 7] reported that the lipophilicity of a group
of organic compounds exhibited a linear relationship with their
narcotic potencies. In 1917, Moore [8] investigated the effects of
chemical “fumigants” on insects and showed that toxicity of these
compounds increased with their boiling point. It can be seen that
the majority of these efforts pertained to the establishment of qual-
itative relationships between chemicals and their respective activity/
property.
Quantitative relationships between chemical structures and
activity/property started to take shape when Hammett [9] intro-
duced a simple equation (later to be known as the Hammett equa-
tion) in 1937 that considers the substituent effect caused by
electron-withdrawing or electron-donating groups as summarized
by the sigma constant, while the rho constant describes the struc-
tural class or chemotype under study. Such an equation allows
determination of electronic effects caused by the placement of sub-
stituent groups at different positions (i.e., ortho-, meta-, and para-)
of the benzene ring. The subsequent work of Taft [10] led to the
introduction of the steric parameter. It is these two contributions
from Hammett and Taft that paved the way for further contribu-
tions from Hansch et al. In the mid-1950s to 1960s, Hansch and
Muir employed the Hammett substituent constant and partition
coefficients in their formulation of equations to correlate the
chemical substituents and growth of Avena plant using plant
growth regulators comprising of phenoxyacetic acids and chloro-
mycetins [11–13]. Finally in 1964, Hansch and Fujita [14] inves-
tigated the biological effects of a wide range of chemicals using a
combination of physicochemical properties as summarized by the
following equation:
1
log = a ⋅ σ + b ⋅π − c ⋅π 2 + d
C
where C is the molar concentration of hormone, σ is the Hammett
parameter, and π is a measure of hydrophobicity. In parallel, Free
and Wilson [15] analyzed such structure-activity relationships using
a binary approach in which the presence or absence of a substituent
AutoWeka 121
was described as 0 and 1, respectively. These landmark reports

marked the beginning of classical QSAR/QSPR. Later in 1969,
Hansch stressed the importance of computers in analyzing struc-
ture-activity relationships [16], which was an observation that is
true to this very day, in which the ever-increasing amount of bio-
logical and chemical data makes computers indispensable in the
analysis of these relationships.
The essence of QSAR/QSPR lies in the predictive ability of
these models to relate a set of explanatory variables (X) with their
response variables (Y). Such explanatory variables of compounds
are represented by molecular descriptors, which are correlated with
functional moieties present in the chemical structure. Descriptors
can be calculated directly from the chemical structure, although in
some circumstances some form of geometry optimization is needed
prior to obtaining the molecular features. Depending on the soft-
ware used, the number of descriptors can span thousands of col-
linear and redundant descriptors, a situation that would typically
call for the use of feature selection. Structure-activity/property
relationships can then be discerned from traditional multiple linear
regression as well as from a wide array of available machine learn-
ing algorithms. The reliability of constructed QSAR/QSPR mod-
els could be assessed from their internal and external validation,
statistical measures of predictive performance, Y-randomization
experiments, etc. Assessment of the applicability domain of the
constructed QSAR/QSPR models may also provide useful knowl-
edge of the range or scope of compounds that are covered by the
model, since the model is only as good as the compounds used to
train it. The aforementioned procedures are common steps in a
typical QSAR/QSPR workflow as summarized in Fig. 1; further
details are discussed in previous review articles [17, 18].
As noted previously, the early days of QSAR/QSPR were pre-
dominantly concerned with its utilization for investigating toxicity
and the narcotic effects of compounds. Over time, this use had
expanded to encompass a wide range of biological activities and
chemical properties, as summarized in Table 1. In a simplistic sys-
tem, compounds are isolated in the sense that they act or behave a
certain way (e.g., as revealed by their boiling point or melting
point) owing directly to their unique chemical structure. To com-
plicate the matter, it may be the case that minor changes to the
chemical structure may exert a drastic influence on its observed
property. This issue is known as the structure-activity cliff and had
been documented previously [19, 20]. The underlying reason for
this is that a majority of biological activities arises from the interac-
tion of compounds with target proteins. A minor change in the
chemical structure, such as the replacement of a methyl group by
an ethyl group, may then cause steric clashes with an amino acid
side chain inside the protein’s binding pocket and thus radically
alter the activity of the compound.
1. Data collection 2. Geometry Optimization
Bioactivity Primary
Database Literature Draw Chemical Initial Geometry Optimization
Structures (AM1)
Conformational Search Refined Geometry Optimization

Compile Data Set (MMFF level) (B3LYP/6-31g(d) )
3. Descriptor Calculation 4. Multivariate Analysis

and Selection
Molecular Descriptors
Optimized Molecular
Structure
Artificial Neural Network Support Vector Machine

Compute Molecular
Descriptors
Parameter Optimization
Linear Polynomial RBF
Filter Redundant Kernel Kernel Kernel
Descriptors Out Number of Hidden Nodes,
Number of Learning Epochs Parameter Optimization
Learning Rate and Momentum
Feature Selection
C, e E, C, e C, g, e
Reduced Subset of Predicted Bioactivity Values
Molecular
Descriptors Predicted Bioactivity Values
Fig. 1 Schematic representation of the QSAR/QSPR workflow
Although QSAR/QSPR may afford a wide range of utility in

the life sciences, there are several potential pitfalls that practitioners
need to be aware of. As this topic is beyond the scope of this chapter,
readers are directed to previously reported accounts of such flaws
[21–25] and solutions [26–29]. In particular, Dearden et al. [30]
discuss 21 common types of error encountered in the QSAR/QSPR
literature; they also provide recommendations on how to tackle
such problems.
Efforts to automate some aspects of the QSAR workflow first
appeared in 2001 when Jewell et al. [31] performed automatic
generation of alignments for 3D-QSAR studies. A subsequent
work from Tervo et al. [32] compared the performance of manual
and automatic alignments. The first automated PLS-based feature
selection was reported in 2004 by Olah et al. [33] in their exhaus-
tive analysis of biological data from WOMBAT. A year later,
Bhonsle et al. [34] reported an automated quasi-4D-QSAR study
of CXCR4 inhibitors based on PLS and scripting language. In the
same year, Cartmell et al. [35] described an automated QSPR
workflow to investigate ADME data sets based on a software archi-
tecture called competitive workflow as implemented in the
Discovery Bus. Zhang et al. [36] introduced the Automated Lazy
AutoWeka 123
Table 1
Types of biological activities modeled by QSAR and selected examples
Type Examples
Absorption Blood-brain barrier penetration
Human intestinal absorption
P-glycoprotein
Skin permeability
Distribution Aqueous solubility (logS)
Octanol-water partition coefficient (logP)
Octanol-water distribution coefficient (logD)
Metabolism Cytochrome P450 induction/inhibition
Excretion Hepatic microsomal intrinsic clearance
Toxicity AMES mutagenicity
Carcinogenicity
Hepatotoxicity
Skin sensitivity
Target/organ cytotoxicity
Receptor agonist/antagonist A3 adenosine receptor antagonist
Angiotensin II receptor antagonist
CCR5 receptor antagonist
Glucagon receptor antagonist
Peroxisome proliferator-activated receptor
gamma agonists
Enzyme activator/inhibitor/modulator Aromatase inhibitor
Dipeptidyl peptidase IV inhibitor
Gamma-aminobutyric acid modulator
Gamma-secretase modulator
Glucokinase activator
Learning QSAR (ALL-QSAR) modeling approach in 2006 as

applied to three sets of anticonvulsant agents. Subsequently in
2007, Bhonsle et al. [37] reported a semiautomated QSAR work-
flow using Tcl-based Cerius2 scripts in their investigation of a set
of insect repellants. Furthermore, Obrezanova et al. [38] described
an automated QSAR modeling approach of ADME data sets using
the Gaussian process. Moreover, Rodgers et al. [39] introduced an
approach to automatically update the QSAR model with measured
data of new compounds in their human plasma protein binding
model. In 2008, in a continuation of their earlier work, Obrezanova
et al. compared the results from automated QSAR models with
those of manual efforts in their study of blood-brain barrier pene-
tration and aqueous solubility. In the same year, Ma et al. [40]
reported an inductive data mining approach in the automatic
generation of decision trees for modeling of the ecotoxicity of
chemicals as well as in the analysis of a historical data from wastewater
treatment plant. In 2011, Wood et al. [41] presented an automated

QSAR procedure employed in AstraZeneca’s AutoQSAR system
as tested against three properties comprising of logD, solubility, and
human plasma protein binding. Furthermore, Sushko et al. intro-
duced the Online Chemical Modeling Environment (OCHEM) as
a web-based tool for automating the process of QSAR modeling.
In 2012, Pérez-Castillo et al. [42] described the genetic-algorithm-
(meta)-ensembles approach for binary classification of five data sets
from the literature. In 2013, Cox et al. [43] reported the QSAR
Workbench, which is based on the Pipeline Pilot workflow tool
and evaluated against two public domain data sets. Furthermore,
Martins and Ferreira [44] introduced software called QSAR mod-
eling for generating and validating QSAR models.
As we can see, development of QSAR models may not be a
straightforward task, especially for the non-bioinformatician,
and this is concomitant with the fact that efforts to automate the
QSAR/QSPR process are an active area of research. These and
many other factors sparked our interests and motivated us to
develop AutoWeka starting from late 2009, which we have been
coding ever since. The project began as an effort to automate our
own data mining workflow as applied to QSAR/QSPR modeling
and finally became publicly and freely available in 2012. The data
mining capabilities, particularly through the use of artificial neural
networks and support vector machines, of AutoWeka are powered
by the popular and widely used Weka machine learning package [45].
It should be noted at the outset that the AutoWeka software is
used for the multivariate analysis phase of QSAR/QSPR modeling,
for which it also performs an extensive parameter optimization, the
results of which could be scrutinized and selected for further
rounds of computation (see Note 1). Nevertheless, specific details
of chemical structure drawing and molecular descriptor generation
are also mentioned in this chapter. It is also worthy of mention that
this chapter will not cover the software development of AutoWeka,
which will be discussed elsewhere.
In the following sections, we will provide an example of the
practical use of AutoWeka in constructing QSAR/QSPR models
(see Note 2). This is demonstrated in a step-by-step manner using
a set of curcumin analogs as a case study.
2 Materials
2.1 Data Source There are several ways in which one can acquire the data that is to
be used for QSAR/QSPR modeling. It can be measured data from
one’s own experiments, compiled from the primary literature,
retrieved from curated databases, or even all of the above. In most
cases, accessibility to the primary literature is heavily dependent on
institutional or personal subscription, while there are a growing
number of open access journals that are freely available for readers.
AutoWeka 125
Typical journals describing the implementation of QSAR/

QSPR models include but are not limited to the following:
●● Bioinformatics
●● Bioorganic & Medicinal Chemistry
●● Bioorganic & Medicinal Chemistry Letters
●● BMC Bioinformatics
●● Briefings in Bioinformatics
●● Chemical Biology & Drug Design
●● Chemometrics and Intelligent Laboratory Systems
●● Chemosphere
●● Computational and Theoretical Chemistry (formerly Journal of
Molecular Structure: THEOCHEM)
●● Computational Biology and Chemistry
●● European Journal of Medicinal Chemistry
●● International Journal of Quantum Chemistry
●● Journal of Chemical Information and Modeling
●● Journal of Cheminformatics
●● Journal of Chemistry
●● Journal of Chemometrics
●● Journal of Computational Biology
●● Journal of Computational Chemistry
●● Journal of Computer-Aided Molecular Design
●● Journal of Enzyme Inhibition and Medicinal Chemistry
●● Journal of Medicinal Chemistry
●● Journal of Molecular Graphics and Modelling
●● Journal of Molecular Modeling
●● Journal of Theoretical and Computational Chemistry
●● Letters in Drug Design & Discovery
●● Medicinal Chemistry Research
●● Molecular Informatics
●● Molecular Simulation
●● PLoS Computational Biology
●● SAR and QSAR in Environmental Research
The following databases contain curated bioactivity data that
have either been deposited by the laboratory generating the data
or compiled from the literature:
●● BindingDB—a database containing ~1,000,000 pieces of
interaction data for ~6,500 protein targets and ~427,000 small
molecules. It is accessible at http://www.bindingdb.org.
●● ChEMBL—a database containing ~12,000,000 pieces of

interaction data for ~9,000 protein targets and ~1,500,000
small molecules. It is accessible at https://www.ebi.ac.uk/
chembldb.
●● DrugBank—a database containing ~6,800 drug entries span-
ning FDA-approved small molecules, FDA-approved peptide/
protein drugs, nutraceuticals, and experimental drugs. It is
accessible at http://www.drugbank.ca.
●● PubChem—is comprised of three linked databases spanning
substance, compound, and bioassay. It is accessible at http://
pubchem.ncbi.nlm.nih.gov.
●● WOMBAT—a database containing ~79,000 in vivo measure-
ment data, ~330,000 compounds, and ~1,900 protein targets.
It is commercially available at http://www.sunsetmolecular.com.
2.2 Drawing Chemical structures can be drawn into the computer using any of
and Refining Chemical the following software tools:
Structures ●● Accelrys Draw (available at http://www.accelrys.com)
●● MarvinSketch (available at http://www.chemaxon.com)
●● OpenEye VIDA (available at http://www.eyesopen.com)
Chemical structure file format conversion tools of use include:
●● Babel (available at http://www.eyesopen.com)
●● Open Babel (available at http://www.openbabel.org)
Molecular structures can be refined using online tools:
●● CORINA Online Demo (available at http://www.molecular-
networks.com/online_demos/corina_demo)
●● COSMOS (accessible at http://cosmos.igb.uci.edu)
In certain cases where the molecular structure is relevant or of
importance for use in prediction, it can also be subjected to geom-
etry optimization via quantum chemical calculations:
●● Gaussian (commercially available at http://www.gaussian.
com)
●● GAMESS (available at http://www.msg.ameslab.gov/gamess)
●● HyperChem (commercially available at http://www.hyper.
com)
●● MOLCAS (commercially available at http://www.molcas.org)
●● MOPAC (available at http://openmopac.net)
2.3 Calculating The next step is to generate numerical descriptions of compounds

Molecular Descriptors to be investigated; a wide range of software is available to perform
this task. For example, common compounds may be available from
AutoWeka 127
the MOLE-db website (Accessible at http://michem.disat.unimib.

it/mole_db). Moreover, quantum chemical descriptors can be
obtained from the aforementioned software such as Gaussian,
GAMESS, etc. Several other categories of molecular descriptors
could be obtained from the following software:
●● CDK Descriptor GUI (available at http://www.rguha.net/
code/java/cdkdesc.html)
●● ChemAxon JChem Calculator Plugins (available at http://
www.chemaxon.com/jchem)
●● Dragon (commercially available from http://www.talete.mi.it)
●● E-Dragon (accessible at http://www.vcclab.org/lab/edragon)
●● MODEL (accessible at http://jing.cz3.nus.edu.sg/cgi-bin/
model/model.cgi)
●● PaDEL (available at http://padel.nus.edu.sg)
2.4 Data Compilation Prior to constructing the QSAR/QSPR model, the block of X
molecular descriptors and Y bioactivity values are combined and
prepared in ARFF file format for use as input to AutoWeka.
●● Spreadsheet program
–– Microsoft Excel (commercially available at http://office.
microsoft.com/excel)
–– OpenOffice (freely available at http://www.openoffice.
org)
●● Text editor
–– Notepad++ (freely available at http://notepad-plus-plus.
org) EditPad Lite (freely available at http://www.edit-
padlite.com)
●● CSV to ARFF converter
–– csv2arff (accessible from http://slavnik.fe.uni-lj.si/markot/
csv2arff/csv2arff.php)
2.5 Multivariate It should be noted here that multivariate analysis or the actual pro-
Analysis cess of constructing the QSAR model is performed using AutoWeka
(Freely available at http://www.mt.mahidol.ac.th/autoweka). It is
in this phase that parameter optimization takes place in an auto-
mated fashion. Completed results will contain information on the
best set of parameters for the selected machine learning algorithm.
2.6 Plotting Graphs After completing AutoWeka calculations, results from parameter
optimization could be visualized using the Python scripts available in
the Download section at http://www.mt.mahidol.ac.th/autoweka.
Alternatively, plots of the results could also be created using statis-
tical graphical software such as SigmaPlot (Commercially available
from http://www.sigmaplot.com).
3 Methods
3.1 Data Source At this phase, the practitioner decides on the data to be used in
their QSAR/QSPR project and therefore retrieves the necessary
information (i.e., chemical name, SMILES, and bioactivity values)
from resources mentioned in Subheading 2.1; the data are then
compiled as outlined in Subheading 3.4. In the case study described
in this chapter, we will be using a data set of 22 curcumin analogs
with DPPH free radical-scavenging activity as previously reported
by Venkateswarlu et al. [46] and modeled by Worachartcheewan
et al. [47]. The bioactivity values were binned to the binary labels
“low” and “high” activity, denoting compounds affording IC50
values of greater than and less than 10 μM, respectively. It should
be noted that in the previously reported QSAR modeling study
described by Worachartcheewan et al. [47], three compounds
(nos. 5, 6, and 10) were identified as outliers and subjected to
removal from the data set, thereby reducing it to 19 compounds.
3.2 Drawing Chemical structures of curcumin analogs (Fig. 2) are drawn into
and Refining Chemical the computer using software described in Subheading 2.2.
Structures Subsequently, chemical structures are either subjected to a rough
energy minimization (using a molecular mechanics force field) or a
more extensive quantum chemical calculation (using a tool such as
HF, B3LYP, etc.) in order to obtain a more refined structure.
These structures can be subjected to an initial geometry refine-
ment using the semiempirical Austin Model 1 (AM1) followed by
Becke’s three-parameter hybrid method using the Lee-Yang-Parr
correlation functional (B3LYP) along with 6-31g(d) basis set as
performed previously by Worachartcheewan et al. [47].
3.3 Calculating Molecular descriptors can be derived from unrefined chemical

Molecular Descriptors structures, which may be applicable for 1D and 2D descriptors.
However, in order to derive 3D descriptors, the chemical struc-
tures must first be subjected to geometry optimization using
quantum chemical calculations as noted above. Depending on
the software used to generate the molecular descriptors, this can
lead to the production of hundreds or thousands of descriptors.
It is common to observe a high degree of collinearity or redun-
dancy among these descriptors, and this issue is best handled by
performing feature selection to select a smaller subset of informa-
tive descriptors for further multivariate analysis. As this issue is
beyond the scope of this chapter, readers are directed to previous
review articles.
3.4 Data Compilation One of the first phases of a QSAR/QSPR study is compiling the
data set of interest, which can be performed by entering essential
data into an appropriate spreadsheet program. For example, a typical
AutoWeka 129
O O O O O O
H3CO OCH3 H3CO
HO
H 1 H
OH H
HO 2 H
OH HO
H 3 H
OH
H
OH O O OH
H H
OH O O H
OH O O
4 5 H3CO 6 OCH3
OCH3 OCH3
O O O O O O
H3CO OCH3 H3CO OCH3
(H3C)2N 7 N(CH3)2 H
HO 8 OHH HO
H 9 H
OH
OCH3 OCH3 OCH3
O O O O O O
Br Br (H3C)3C C(CH3)3 (H3C)3C C(CH3)3
H
HO 10 OHH HO
H 11 OH
H H
HO 12 OHH
OCH3 OCH3 C(CH3)3 C(CH3)3
O O O O O O
H3COOC COOCH3 HOOC COOH H
HO OH
H
H
HO 13 H
OH HO
H 14 H
OH HO
H 15 H
OH
O O H
OH O O OH
H O O
HO
H OCH3 H
OH
H
HO 16 OH
H 17 H
HO 18 OH
H
OH
H H
OH
O O O O O O
HO
H H
OH H
HO H
OH H
HO OH
H
HO
H 19 OHH HO
H 20 OH
H HO
H 21 OH
H
OCH3 OCH3 OCH3 H
OH
O O
H
HO OH
H
H
HO 22 OH
H
OH
H OH
H
Fig. 2 Chemical structures of 22 curcumin analogs
molecule or compound will occupy a row in the spreadsheet where

the columns can contain the following information: compound
name, compound ID, XC50 of bioactivity (where X refers to the
type of bioactivity such as inhibition concentration, effective con-
centration, etc.), SMILES notation, molecular descriptors (molec-
ular weight, logP, number of hydrogen bond donor atoms, number
of hydrogen bond acceptor atoms, etc.), and the reference or
data source from which the compounds were obtained (Fig. 3).
Therefore, as there are 19 curcumin analogs (after removal of three
outliers) for this case study, this would correspond to 19 rows plus
the header row (containing the descriptor labels of each column)
resulting in a total of 20 rows.
The next step is to transform our data set from the spreadsheet
format to an ARFF file format that is compatible with Weka, since
this is the machine learning package that powers the software
developed in AutoWeka. This spreadsheet to ARFF transformation
can be carried out using a text editor or alternatively using the
online csv2arff tool. Examples of ARFF input files are shown in
Fig. 3 Screenshot of compiled data using Microsoft Excel
Tables 2 and 3 with quantitative and qualitative Y labels. Line 1

represents the name of the data set, lines 3–8 represent the descrip-
tor names, and lines 11–29 represent the block of descriptors and
their Y values or labels as correspondingly specified by the follow-
ing terms with the @ symbol preceding it: RELATION,
ATTRIBUTE, and DATA. The NUMERIC terms that follow the
descriptor names in lines 3–8 are syntax that are recognized by the
Weka program as descriptors having quantitative values, whereas
the braces {} encapsulating the Low and High terms are also syntax
that are recognized by the Weka program as descriptors having
qualitative values. It should be noted that the Y descriptor is typi-
cally located as the last variable, whereas X descriptors precede it.
Generally, QSAR modeling of data sets having quantitative or
qualitative Y variables is subjected to either regression or classifica-
tion analysis, respectively.
3.5 Machine Before we move on to the multivariate analysis phase and actually
Learning Algorithms construct the QSAR model, it is pertinent to first provide a glimpse
of the algorithmic details of machine learning algorithms that
are commonly used in QSAR/QSPR on biological [47–56] and
chemical systems [57–61] and which are implemented in AutoWeka.
3.5.1 Artificial Neural The artificial neural network (ANN) is a well-known multivariate
Network method that is commonly used to develop QSAR/QSPR models.
ANN takes its inspiration from the biological brain with its organi-
zation of neurons [62]. The single-layer perceptron (SLP), which
AutoWeka 131
Table 2
Example of ARFF input file of curcumin analogs with
quantitative Y variable
@RELATION RadicalScavengingRegression
@ATTRIBUTE dipole NUMERIC

@ATTRIBUTE gap NUMERIC
@ATTRIBUTE hardness NUMERIC
@ATTRIBUTE softness NUMERIC
@ATTRIBUTE OH NUMERIC
@ATTRIBUTE pIC50 NUMERIC
@DATA
4.82,−0.134,0.067,7.463,2,−1.322
3.53,−0.141,0.071,7.092,2,−1.531
3.793,−0.143,0.072,6.993,2,−1.519
5.91,−0.124,0.062,8.065,2,−1.38
5.819,−0.127,0.064,7.874,0,−1.716
7.447,−0.133,0.067,7.519,2,−1.415
6.721,−0.129,0.065,7.752,2,−1.407
4.584,−0.142,0.071,7.042,2,−1.681
5.27,−0.139,0.07,7.194,2,−1.633
5.122,−0.142,0.071,7.042,2,−2
3.643,−0.146,0.073,6.849,2,−2
3.277,−0.134,0.067,7.463,4,−0.778
3.949,−0.134,0.067,7.463,3,−0.845
6.099,−0.125,0.063,8,4,−0.903
2.758,−0.138,0.069,7.246,3,−0.881
1.995,−0.129,0.065,7.752,4,−0.732
3.507,−0.127,0.064,7.874,4,−0.799
3.901,−0.129,0.065,7.752,5,−0.716
4.116,−0.127,0.064,7.874,6,−0.663
Table 3
Example of ARFF input file of curcumin
analogs with qualitative Y variable
@RELATION
RadicalScavengingClassification
@ATTRIBUTE dipole NUMERIC

@ATTRIBUTE gap NUMERIC
@ATTRIBUTE hardness NUMERIC
@ATTRIBUTE softness NUMERIC
@ATTRIBUTE OH NUMERIC
@ATTRIBUTE class {Low, High}
@DATA
4.82,−0.134,0.067,7.463,2,Low
3.53,−0.141,0.071,7.092,2,Low
3.793,−0.143,0.072,6.993,2,Low
5.91,−0.124,0.062,8.065,2,Low
5.819,−0.127,0.064,7.874,0,Low
7.447,−0.133,0.067,7.519,2,Low
6.721,−0.129,0.065,7.752,2,Low
4.584,−0.142,0.071,7.042,2,Low
5.27,−0.139,0.07,7.194,2,Low
5.122,−0.142,0.071,7.042,2,Low
3.643,−0.146,0.073,6.849,2,Low
3.277,−0.134,0.067,7.463,4,High
3.949,−0.134,0.067,7.463,3,High
6.099,−0.125,0.063,8,4,High
2.758,−0.138,0.069,7.246,3,High
1.995,−0.129,0.065,7.752,4,High
3.507,−0.127,0.064,7.874,4,High
3.901,−0.129,0.065,7.752,5,High
4.116,−0.127,0.064,7.874,6,High
AutoWeka 133
Input Output
layer layer
x1
x2 y1
x3
Fig. 4 Schematic architecture of the single-layer perceptron
is the classical model of ANN, is the simplest type of ANN; it is

comprised of several input nodes and a single output node as
shown in Fig. 4.
For a given training sample D = (x1, y1), (x2, y2), …, (xN, yN), we
can estimate the variable yi by combining the weighted sum of its
N inputs as follows:
N 
y i = θ  ∑wi xi  (1)
 i =1 
where yi belongs to class 1 if the weighted sum is greater than the

selected threshold; otherwise, yi belongs to class 0. θ(•) is the acti-
vation function that maps the weighted sum of inputs to the out-
put. The most popularly currently used activation functions are
the logistic sigmoid (1 / 1 + e − xi ) and hyperbolic tangent (tanh(xi)).
In practical application, SLP is unable to solve a nonlinearly sepa-
rable data problem, owing to the fact that during the learning pro-
cess, this approach cannot properly predict all data points on D.
Thus, the multilayer perceptron (MLP) was proposed for handling
nonlinearly separable data by adding one or more hidden layers
(see Fig. 5) [63, 64].
The MLP approach is a type of supervised learning method in
which the back-propagation algorithm is applied to estimate the
optimized parameter wi by changing the weighted connection,
which is dependent on the magnitude of the error (the difference
between the actual and predicted value). The back-propagation
algorithm is comprised of two major phases: (1) forward phase and
(2) backward phase, as briefly described below.
Step 1. Initialize a connection weight wi with a small random value.
Step 2. Randomly select a training sample Di ⊂ D, |Di| = p where
p < N.
Input Hidden Output

layer layer layer
x1 h1
x2 h2 y
x3 h3
Fig. 5 Schematic architecture of the multilayer perceptron
Step 3. Calculate the partial derivative of weight wi.

Step 4. Update weight wi according to the following equations:
∂E (n )
wij (n + 1) = wij (n ) + η (2)
∂wij

∂E
wij (n + 1) = wij (n ) + η (3)
∂wij

where wij(n) is the value of wij prior to updating by n times while
wij(n + 1) is the value of wij after such updates and η is the learning
rate that determines both the convergence rate and stability of the
training process while E is the cost or error function.
Step 5. Repeat steps 2–4 for every training sample Di, and repeat
for these sets until the cost of output errors is minimized.
3.5.2 Support Vector The support vector machine (SVM) is a popular learning method
Machine that had been adopted for solving a plethora of problems via clas-
sification and regression analysis. A notable property of SVM is its
estimation of model parameters by means of the convex
optimization approach that guarantees that a local solution is also
the global optimum [65]. Vapnik first introduced this technique
based on the principles of structure risk minimization of the statis-
tical learning theory [66, 67]. In practice, the SVM method
constructs a maximum-margin hyperplane to separate two classes.
To solve nonlinearly separable data, the SVM method acts in con-
junction with a mapping function Φ(x) : x ∈ ℜM → ℜP that is used to
transform the original data set of M-dimension onto a higher
dimensional space or feature space of P-dimension where M ≪ P.
Subsequently, a simple linear classifier f(xi) can then be used for
classifying P-dimensional samples [68, 69] (shown in Fig. 6).
AutoWeka 135
Input space Feature space
Kernel function
Φ(x)
Fig. 6 Schematic representation of relationship between input and feature spaces using a mapping function
The kernel function K(xi, xj) is used to represent the mapping

function by taking the inner product between two samples xi and
xj in D, which is defined as:
N
( )
K x i ,x j = Φ ( x i ) Φ x j =
T
( ) ∑Φ (x ) Φ (x ) i
T
j (4)
i , j =1

In practice, given D, SVM is used to construct a linear function
f(xi) representing the correlation between the structure and bio-
logical activities/chemical properties of data xi:
N
f ( xi ) = ∑wi Φ ( xi ) + b (5)
i =1
where wi ∈ ℜ are the coefficients, b ∈ ℜ is the bias, and N is the
M
number of samples. The most popularly used kernel comprises the

following:
●● Linear kernel:

Φ ( xi ) Φ x j
T
( ) (6)
●● Polynomial kernel:
(1 + Φ ( x ) Φ ( x ) )
T d
i j (7)

where d = 2, 3, and 4 (it should be noted that d = 1 for a linear
kernel).
●● Radial basis function (RBF) kernel:

( (
exp − γ xi − x j )) (8)
where γ is greater than 0.

Table 4
Summary of computational methods used in QSAR/QSPR modeling
Built-in
feature Single/
Method Advantage Disadvantage selector ensemble
ANN Performs well on complex data Low interpretability No Single
SVM Solves nonlinearly separable data Low interpretability No Single
PCA Summarizes a data set without Does not consider relationship Yes Single
losing too much variation between X and Y
PLS Simple and interpretable model Requires cross-term Yes Single
DT Simple and interpretable model Requires a number of training Yes Single
data sets
RF High interpretability and low Complexity method Yes Ensemble
risk of over-fitting
ANN, SVM, PCA, PLS, DT, and RF are acronyms for artificial neural network, support vector machine, principal com-
ponent analysis, partial least squares regression, decision tree, and random forest, respectively
In regression problems, when y is a numerical value, estimation

of the parameter wi can be achieved by utilizing the ε-insensitive
loss function (Lε(y, f(x, w))) [70, 71] described as follows:
| y − f ( x ,w ) | −ε, y − f ( x ,w ) |≥ ε
L ε ( y , f ( x ,w ) ) =  (9)
 0, y − f ( x ,w ) |< ε

where y is the actual value, f(x, w) is the predicted function for esti-
mating the output value, and ε is the insensitivity parameter.
Although SVM is widely popular, SVM and ANN methods both
suffer from the fact that they are a black-box approach and there-
fore lack interpretability; they do not readily indicate which
feature(s) is of greatest importance in the structure of the predic-
tion model. Table 4 highlights and compares the advantages and
disadvantages of SVM and ANN in the context of other commonly
used learning algorithms in QSAR/QSPR modeling.
3.6 Multivariate We will now proceed with the step-by-step case study of construct-
Analysis ing the QSAR model using AutoWeka. Although we try to provide
as much detail as possible on the practical aspects of using AutoWeka,
some adjustments may need to be made as the reader adapts this
protocol to their own projects. As previously mentioned in
Subheading 3.1, the case study used in this chapter is based on the
set of 22 curcumin analogs reported by Venkateswarlu et al. [46].
Let us start by going to the website of AutoWeka that is
accessible at http://www.mt.mahidol.ac.th/autoweka. Readers
can click on either the Download link on top or the orange
Download button down below (Fig. 7).
AutoWeka 137
Fig. 7 Screenshot of AutoWeka’s website
On the following page, readers have access to the AutoWeka

User Manual, the Python scripts for creating graphs and plots, as
well as the AutoWeka software. After agreeing to the terms of the
license agreement, click on the “I Agree, Please Proceed to Download
AutoWeka” button (Fig. 8) to proceed. The next page will then
bring out a registration form that users can fill out, and after its
submission, the Download link will appear. The zip file of the soft-
ware is approximately 19 MB in size. After successfully download-
ing the file, unzip it to a desired location, and the following
contents as shown in Fig. 9 (left panel) will appear. After double-
clicking on the AutoWeka.exe file (left panel) the program window
will appear (right panel) as shown in Fig. 9. The menu bar contains
three possible options, Run, Tools, and About, which correspond-
ingly allow users to run the machine learning algorithms for con-
structing the QSAR/QSPR model, adjust the memory value to
use (Fig. 10), and provide access to the About window.
Now that the software is up and running, let us get started
with setting up an ANN calculation (Fig. 11) by first clicking on
Run → Artificial Neural Network. This brings up a new window
that allows us to set up the various ANN parameters or choose the
easiest option, which is to use the default parameters. Here we
click on the Browse button and select the ARFF input file of inter-
est (in our case, the Curcumin_regression.arff) file. Next, click on
the Default button and finally the Start button to proceed with
building the ANN model.
Fig. 8 Screenshot of the Download page on AutoWeka’s website
Subsequently, a new window appears that summarizes the

parameter settings that will be used for the calculation. Upon click-
ing on the OK button, a pop-up window asks for confirmation of
the name of an automatically generated folder (Fig. 12). Here, we
can click on the OK button again to use the default name. Next,
the progress window (Fig. 13) appears; the time required for
completion will depend on the complexity of the input data. Upon
completion, a pop-up notification box appears, and we can then
click on the OK button to finalize the calculation.
All calculation files are located in the Results folder (Fig. 14) that
also goes by the same folder name in the root folder of AutoWeka,
meaning that if we unpacked AutoWeka directly to the C drive,
the relative path of the root folder would be C:\AutoWeka\ and
thereby the Results folder could be found at C:\AutoWeka\Results\.
Inside the Results folder, double-click on the Curcumin_analogs
AutoWeka 139
Fig. 9 Screenshot of the contents of the zip file of AutoWeka software (left panel) and upon opening the
AutoWeka program (right panel)
folder to show a sub-folder that is automatically named according to

the learning method used and the selected parameters. For example,
if ANN was selected as the learning method and hidden nodes of
1–25 was chosen then the folder name would start with ANN_
Hidden_1_to_25. The same convention also applies to the other
parameters that will be appended at the end of the abovementioned
folder name.
Inside the ANN results folder (Fig. 15), there are three
sub-folders corresponding to the three ANN parameters, accord-
ingly named as HiddenNode, LearningAndMomentum, and
TrainingTime. Double-clicking on one of the sub-folders reveals a
collection of sequentially numbered files where one parameter set-
ting will generate one calculation output file. Thus, for an investi-
gation of 25 hidden nodes (also bearing in mind that for each
parameter investigated, ten separate runs are performed owing to
the inherently random nature of the weight initialization of the
back-propagation algorithm of ANN), a total of 25 × 10 = 250 out-
put files will be generated. Subsequently, an average value for each
investigated parameter will be derived from these ten calculations
and saved into a new file called AvgHidden.txt. As data values
within the AvgHidden.txt results file are in a tab-delimited format,
Fig. 10 Screenshot of adjusting the memory settings to be used by AutoWeka
Fig. 11 Screenshot of setting up an ANN calculation

AutoWeka 141
Fig. 12 Screenshot of pop-up windows that appear prior to ANN calculations
Fig. 13 Screenshot of an ongoing (left panel) and completed (right panel) ANN calculation
Fig. 14 Screenshot of the method for accessing the calculation results folder
it can be readily visualized by importing the file (or copying and

pasting the data) into Microsoft Excel (Fig. 16). For convenience,
the best parameter settings along with a summary of the statistical
performance are provided in the SummaryHidden.txt file.
An alternative way of assessing the raw numerical data of the
calculation results is to make a visual representation of it by creat-
ing graphical plots. This can be carried out by using the prewritten
Python scripts or the graphical plot software mentioned in
Subheading 2.6. Graphical plots created by the former are shown
in Fig. 17.
Fig. 15 Screenshot of calculation results as divided by the three folders of the optimized parameters:
HiddenNode, LearningRateAndMomentum, as well as TrainingTime. Shown are contents from the AvgHidden.
txt file found in the HiddenNode folder. It should be noted that the same structure and organization of calcula-
tion results apply to the other two folders
Fig. 16 Screenshot of AvgHidden.txt file as depicted in Microsoft Excel

Fig. 17 Graphical plots from the parameter optimization process for (a) the number
of hidden nodes, (b) the number of learning epochs, and (c) the learning rate and
momentum
AutoWeka 145
4 Notes
1. AutoWeka can be used in conjunction with Weka if users would

like to benchmark their models with other learning algorithms.
2. In a typical QSAR modeling project, the rate-limiting step,
that is, the lengthiest step, is generally that of data compilation
and curation. It is here that critical errors (e.g., correctness of
data entry) must be identified to ensure data integrity. The
next most time-consuming step is the parameter optimization
phase, in which several iterations of parameter fine-tuning are
carried out to produce the best performance. AutoWeka han-
dles the latter point automatically as it seeks the best set of
parameters.
Acknowledgments
This work was supported by Mahidol University via the Goal-

Oriented Research Grant to C.N.; postdoctoral fellowship to W.S.;
research assistantships to P.M., S.J., and L.P.; and partial financial
support to S.S.
References
1. Brodin A (1858) On the analogy of arsenic and 10. Taft RW (1952) Polar and steric substituent
phosphoric acid with respect to chemical and constants for aliphatic and o-benzoate groups
toxicology. Medico-Surgical Academy, St. from rates of esterification and hydrolysis of
Petersburg, Russia esters1. J Am Chem Soc 74(12):3120–3128
2. Cros A (1863) Action de l’alcool amylique sur 11. Hansch C, Maloney PP, Fujita T et al (1962)
l’organisme. University of Strasbourg, Strasbourg Correlation of biological activity of phenoxyace-
3. Kekulé A (1865) Sur la constitution des sub- tic acids with Hammett substituent constants
stances aromatiques. Bull Soc Chim Fr 3:98 and partition coefficients. Nature 194:178–180
4. Richardson B (1869) Physiological research on 12. Hansch C, Muir RM, Fujita T et al (1963) The
alcohols. Med Times Gaz 2:703–706 correlation of biological activity of plant growth
5. Richet C (1893) On the relationship between regulators and chloromycetin derivatives with
the toxicity and the physical properties of sub- Hammett constants and partition coefficients.
stances. Compt Rendus Seances Soc Biol J Am Chem Soc 85(18):2817–2824
9:775–776 13. Hansch C, Muir RM (1950) The ortho effect
6. Overton E (1897) Osmotic properties of cells in plant growth-regulators. Plant Physiol
in the bearing on toxicology and pharmacol- 25(3):389
ogy. Z Phys Chem 22:189–209 14. Hansch C, Fujita T (1964) p-σ-π analysis. A
7. Meyer H (1899) On the theory of alcohol nar- method for the correlation of biological activity
cosis. Arch Exp Pathol Pharmacol 42:109–118 and chemical structure. J Am Chem Soc 86(8):
8. Moore W (1917) Volatility of organic com- 1616–1626
pounds as an index of the toxicity of their 15. Free SM Jr, Wilson JW (1964) A mathematical
vapors to insects. J Agric Res 10(7):365 contribution to structure-activity studies. J Med
9. Hammett LP (1937) The effect of structure Chem 7:395–399
upon the reactions of organic compounds. 16. Hansch C (1969) Quantitative approach to
Benzene derivatives. J Am Chem Soc 59(1): biochemical structure-activity relationships. Acc
96–103 Chem Res 2(8):232–239
17. Nantasenamat C, Isarankura-Na-Ayudhya C, between 3D QSAR models obtained from

Naenna T et al (2009) A practical overview manual and automated alignment. J Chem Inf
of quantitative structure-activity relationship. Comput Sci 44(3):807–816
Excli J 8:74–88 33. Olah M, Bologa C, Oprea TI (2004) An auto-

18. Nantasenamat C, Isarankura-Na-Ayudhya C, mated PLS search for biologically relevant
Prachayasittikul V (2010) Advances in compu- QSAR descriptors. J Comput Aided Mol Des
tational methods to predict the biological activ- 18(7–9):437–449
ity of compounds. Expert Opin Drug Discov 34. Bhonsle JB, Wang Z-X, Tamamura H et al
5(7):633–654 (2005) A simple, automated quasi-4D-QSAR,
19. Medina-Franco JL, Martinez-Mayorga K, quasi-multi way PLS approach to develop
Bender A et al (2009) Characterization of highly predictive QSAR models for highly
activity landscapes using 2D and 3D similarity flexible CXCR4 inhibitor cyclic pentapeptide
methods: consensus activity cliffs. J Chem Inf ligands using scripted common molecular
Model 49(2):477–491 modeling tools. QSAR Comb Sci 24(5):
20. Bajorath J (2012) Modeling of activity land- 620–630
scapes for drug discovery. Expert Opin Drug 35. Cartmell J, Enoch S, Krstajic D et al (2005)
Discov 7(6):463–473 Automated QSPR through competitive work-
21. Doweyko AM (2008) QSAR: dead or alive? J flow. J Comput Aided Mol Des 19(11):
Comput Aided Mol Des 22(2):81–89 821–833
22. Doweyko AM (2008) Is QSAR relevant to 36. Zhang S, Golbraikh A, Oloff S et al (2006) A
drug discovery? IDrugs 11(12):894–899 novel automated lazy learning QSAR (ALL-
23. Tropsha A, Golbraikh A (2007) Predictive QSAR) approach: method development, appli-
QSAR modeling workflow, model applicability cations, and virtual screening of chemical
domains, and virtual screening. Curr Pharm Des databases using validated ALL-QSAR models.
13(34):3494–3504 J Chem Inf Model 46(5):1984–1995
24. Golbraikh A, Tropsha A (2002) Beware of q2! 37. Bhonsle JB, Bhattacharjee AK, Gupta RK

J Mol Graph Model 20(4):269–276 (2007) Novel semi-automated methodology
for developing highly predictive QSAR models:
25. Huang J, Fan X (2011) Why QSAR fails: an application for development of QSAR mod-
empirical evaluation using conventional compu- els for insect repellent amides. J Mol Model
tational approach. Mol Pharm 8(2):600–608 13(1):179–208
26. Tropsha A, Gramatica P, Gombar VK (2003) 38. Obrezanova O, Csanyi G, Gola JM et al (2007)
The importance of being earnest: validation is Gaussian processes: a method for automatic
the absolute essential for successful application QSAR modeling of ADME properties. J Chem
and interpretation of QSPR models. QSAR Inf Model 47(5):1847–1857
Comb Sci 22(1):69–77
39. Rodgers SL, Davis AM, Tomkinson NP et al
27. Tropsha A (2010) Best practices for QSAR (2007) QSAR modeling using automatically
model development, validation, and exploita- updating correction libraries: application to a
tion. Mol Inf 29(6–7):476–488 human plasma protein binding model. J Chem
28. Fourches D, Muratov E, Tropsha A (2010) Inf Model 47(6):2401–2407
Trust, but verify: on the importance of chemi- 40. Ma CY, Buontempo FV, Wang XZ (2008)
cal structure curation in cheminformatics and Inductive data mining: automatic generation of
QSAR modeling research. J Chem Inf Model decision trees from data for QSAR modelling
50(7):1189–1204 and process historical data analysis. Comput
29. Scior T, Bender A, Tresadern G et al (2012) Aid Chem Eng 25:581–586
Recognizing pitfalls in virtual screening: a criti- 41. Wood DJ, Buttar D, Cumming JG et al (2011)
cal review. J Chem Inf Model 52(4):867–881 Automated QSAR with a hierarchy of global
30. Dearden JC, Cronin MT, Kaiser KL (2009) and local models. Mol Inf 30(11–12):960–972
How not to develop a quantitative structure- 42. Perez-Castillo Y, Lazar C, Taminau J et al
activity or structure-property relationship (2012) GA(M)E-QSAR: a novel, fully automatic
(QSAR/QSPR). SAR QSAR Environ Res genetic-algorithm-(meta)-ensembles approach
20(3–4):241–266 for binary classification in ligand-
based drug
31. Jewell NE, Turner DB, Willett P et al (2001) design. J Chem Inf Model 52(9):2366–2386
Automatic generation of alignments for 3D 43. Cox R, Green DV, Luscombe CN et al (2013)
QSAR analyses. J Mol Graph Model 20(2): QSAR workbench: automating QSAR modeling
111–121 to drive compound design. J Comput Aided
32. Tervo AJ, Nyronen TH, Ronkko T et al (2004) Mol Des 27(4):321–336
Comparing the quality and predictiveness
AutoWeka 147
44. Martins JPA, Ferreira MMC (2013) QSAR chemical properties of templates from
modeling: a new open source computational molecular imprinting literature using interac-
package to generate and validate QSAR models. tive text mining approach. Chemometr Intell
Quim Nova 26:554–560 Lab Syst 116:128–136
45. Hall M, Frank E, Holmes G et al (2009) The 58. Nantasenamat C, Isarankura-Na-Ayudhya C,
WEKA data mining software: an update. Tansila N et al (2007) Prediction of GFP spec-
SIGKDD Explorations 11 (1) tral properties using artificial neural network. J
46. Venkateswarlu S, Ramachandra MS, Subbaraju Comput Chem 28(7):1275–1289
GV (2005) Synthesis and biological evaluation 59. Nantasenamat C, Naenna T, Isarankura N-AC
of polyhydroxycurcuminoids. Bioorg Med et al (2005) Quantitative prediction of imprint-
Chem 13(23):6374–6380 ing factor of molecularly imprinted polymers
47. Worachartcheewan A, Nantasenamat C, by artificial neural network. J Comput Aid Mol
Isarankura-Na-Ayudhya C et al (2011) Des 19(7):509–524
Predicting the free radical scavenging activity 60. Nantasenamat C, Isarankura-Na-Ayudhya C,
of curcumin derivatives. Chemometr Intell Lab Naenna T et al (2007) Quantitative structure-
Syst 109(2):207–216 imprinting factor relationship of molecularly
48. Mandi P, Nantasenamat C, Srungboonmee K imprinted polymers. Biosens Bioelectron
et al (2012) QSAR study of anti-prion activity 22(12):3309–3317
of 2-aminothiazoles. Excli J 11:453–467 61. Nantasenamat C, Srungboonmee K, Jamsak S
49. Nantasenamat C, Isarankura-Na-Ayudhya C, et al (2013) Quantitative structure-property
Naenna T et al (2008) Prediction of bond dis- relationship study of spectral properties of
sociation enthalpy of antioxidant phenols by green fluorescent protein with support vector
support vector machine. J Mol Graph Model machine. Chemometr Intell Lab Syst 120:
27(2):188–196 42–52
50. Nantasenamat C, Li H, Mandi P et al (2013) 6 2. McCulloch WS, Pitts W (1943) A logical calcu-
Exploring the chemical space of aromatase lus of the ideas immanent in nervous activity.
inhibitors. Mol Div. doi:10.1007/s11030- Bull Math Biophys 5(4):115–133
11013-19462-x 63. Lawrence J (1993) Introduction to neural net-
51. Nantasenamat C, Piacham T, Tantimongcolwat works: design, theory, and applications, 6th
T et al (2008) QSAR model of the quorum- edn. California Scientific Software, California
quenching N-acyl-homoserine lactone lac- 64. Smith M (1993) Neural networks for statisti-
tonase activity. J Biol Syst 16(2):279–293 cal modeling. Van Nostrand Reinhold,
52. Pingaew R, Tongraung P, Worachartcheewan A New York
et al (2012) Cytotoxicity and QSAR study of (thio) 65. Bishop CM, Nasrabadi NM (2006) Pattern
ureas derived from phenylalkylamines and pyri- recognition and machine learning, vol 1.
dylalkylamines. Med Chem Res 22:4016-4029 Springer, New York
53. Prachayasittikul S, Wongsawatkul O, 66. Vapnik V (2000) The nature of statistical learn-
Worachartcheewan A et al (2010) Elucidating the ing theory. Springer, New York
structure-activity relationships of the vasorelax- 67. Vapnik V (1998) Statistical learning theory.
ation and antioxidation properties of thionico- Wiley, New York
tinic acid derivatives. Molecules 15(1):198–214 68. Cortes C, Vapnik V (1995) Support-vector
54. Thippakorn C, Suksrichavalit T, Nantasenamat networks. Mach Learn 20(3):273–297
C et al (2009) Modeling the LPS neutralization 69. Cristianini N, Shawe-Taylor J (2000) An intro-
activity of anti-endotoxins. Molecules 14(5): duction to support vector machines and other
1869–1888 kernel-based learning methods. Cambridge
55. Worachartcheewan A, Nantasenamat C, University Press, Cambridge
Isarankura-Na-Ayudhya C et al (2013) Predicting 70. Platt JC (1999) Fast training of support vector
antimicrobial activities of benzimidazole deriva- machines using sequential minimal optimiza-
tives. Med Chem Res 22:5418–5430 tion. In: Schoelkopf B, Burges C, Smola A
56. Worachartcheewan A, Nantasenamat C, (eds) Advances in kernel methods: support vec-
Naenna T et al (2009) Modeling the activity of tor learning. MIT Press, Cambridge, USA,
furin inhibitors using artificial neural network. pp 185–208
Eur J Med Chem 44(4):1664–1673 71. Smola AJ, Schölkopf B (2004) A tutorial on
57. Nantasenamat C, Li H, Isarankura-Na- support vector regression. Stat Comput 14(3):
Ayudhya C et al (2012) Exploring the physico- 199–222
Chapter 9
Ligand Biological Activity Predictions Using Fingerprint-

Based Artificial Neural Networks (FANN-QSAR)
Kyaw Z. Myint and Xiang-Qun Xie
Abstract
This chapter focuses on the fingerprint-based artificial neural networks QSAR (FANN-QSAR) approach to
predict biological activities of structurally diverse compounds. Three types of fingerprints, namely ECFP6,
FP2, and MACCS, were used as inputs to train the FANN-QSAR models. The results were benchmarked
against known 2D and 3D QSAR methods, and the derived models were used to predict cannabinoid (CB)
ligand binding activities as a case study. In addition, the FANN-QSAR model was used as a virtual screen-
ing tool to search a large NCI compound database for lead cannabinoid compounds. We discovered several
compounds with good CB2 binding affinities ranging from 6.70 nM to 3.75 μM. The studies proved that
the FANN-QSAR method is a useful approach to predict bioactivities or properties of ligands and to find
novel lead compounds for drug discovery research.
Key words Quantitative structure-activity relationship (QSAR), Fingerprint, Artificial neural networks
(ANN), Biological activity, Cannabinoid
1 Introduction
Quantitative structure-activity relationship (QSAR) studies play

essential roles in pharmaceutical research to identify and generate
high-quality leads in the early stages of drug discovery [1–3].
QSAR studies help reduce the costly failures of drug candidates by
identifying promising lead compounds and reducing the number
of costly experiments. Such studies correlate chemical or structural
features of compounds with their bioactivities. Descriptors are
used to encode molecular features, and a mathematical relationship
among a set of descriptors and biological endpoints of existing
compounds is derived to construct a QSAR model which is then
used to predict activities of new structures. Hence it is important
to have a robust model which can learn important molecular
features from a set of training compounds and then recognize such
features in new structures to effectively predict their biological
endpoints or activities.
149
150 Kyaw Z. Myint and Xiang-Qun Xie
The artificial neural networks (ANN) method is a machine

learning algorithm with adaptive learning behavior in which the
algorithm learns from previous examples and adapts to changes in
input parameters. In comparison with receptor docking search [4],
ligand similarity search [5, 6], machine-learning-based search
[7, 8], and QSAR approaches [9–11], the ANN method possesses
good generalization and pattern recognition ability for unseen
data. Such features make it effective and robust for nonlinear
regression problems with multiple inputs. In fact, several studies
have used the ANN algorithm to predict molecular properties or
biological endpoints of chemical analogs in several case studies
such as anti-diabetes, anticancer, anti-HIV, and allergenicity pre-
diction [12–16].
In this chapter, we focus on the fingerprint-based ANN-QSAR
(FANN-QSAR) research work [17] in predicting biological activi-
ties of structurally diverse cannabinoid (CB) ligands using ANN.
To the best of our knowledge, there have been no previous studies
which have used molecular fingerprints as descriptors to predict
biological activities (such as pIC50 or pKi), although a few studies
have been reported to predict ligand classes [18, 19]. Three types
of molecular fingerprints were used as network inputs to train
ANN-QSAR models, and the results were compared to well-known
2D and 3D QSAR methods [1] using five data sets. As a case study,
we used the FANN-QSAR method to predict binding affinities
of cannabinoid ligands using a large and structurally diverse CB
ligand data set [20]. In addition, we applied the method as a vir-
tual screening tool to find new cannabinoid ligands from a large
NCI database containing over 200,000 compounds, and we found
three compounds with good cannabinoid receptor binding affini-
ties which were experimentally validated. The results demonstrated
that combination of molecular fingerprints and ANN can lead to a
reliable and robust high-throughput virtual screening method
which can be a useful tool in chemogenomics and computer-aided
drug discovery research.
2 Methods
2.1 Data Sets As shown in Table 1, a total of six data sets were used in this study.
Five of them were compiled by Sutherland et al. [21] and were
downloaded from their supplemental data. The sixth data set, can-
nabinoid receptor 2 (CB2) data set, was curated by the Xie lab [5,
20]. For this data set, if there were more than one reported CB2
activity for a ligand, an average activity was used.
Once the structure-activity relationship (SAR) data were col-
lected, each data set was then divided into training, validation,
and testing sets. For the ACE, AchE, BZR, COX2, and DHFR
data sets, the same training and testing data sets provided by
Ligand Biological Activity Predictions 151
Table 1
Summary of six data sets used in the FANN-QSAR model development
Target/receptor name Number of inhibitors pIC50 range Reference

Angiotensin-converting enzyme (ACE) 114 2.1–9.9 [45]
Acetylcholinesterase (AchE) 111 4.3–9.5 [46, 47]
Benzodiazepine receptor (BZR) 147 5.5–8.9 [48]
Cyclooxygenase-2 (COX2) 282 4.1–9.0 [49–58]
Dihydrofolate reductase inhibitors (DHFR) 361 3.3–9.8 [59–63]
Cannabinoid receptor subtype-2 (CB2) 1,699 3.9–10.8 [20]
Table 2
Number of training, validation, and testing set compounds in each data set
ACE AchE BZR COX2 DHFR CB2

Training set 69 67 89 170 214 1,361
Validation set 7 7 9 18 23 169
Test set 38 37 49 94 124 169
Total 114 111 147 282 361 1,699
Sutherland et al. were used in order to allow the direct comparison

of FANN-QSAR models to 3D and 2D QSAR models reported
[21]. For each data set, 10 % of randomly selected compounds
from the training set were used as a validation set. For the canna-
binoid data set, the training and test sets were randomly divided.
The training set contained 80 % of the compounds while the test
set contained 10 %; the other 10 % were used as a validation set.
The training set was used to train the model while the validation
set was used to prevent overfitting of the model. The test set was
used as an external set to evaluate the generalization ability of the
trained FANN-QSAR models. The numbers of compounds found
in each training, validation, and test sets for each data set are sum-
marized in Table 2. For statistical modeling, the process was
repeated five times for each data set, resulting in five different
pairs of randomly divided training and test sets.
2.2 Implementation A feed-forward back-propagation neural network method was

of Fingerprint-Based implemented using MATLAB® R2007b Neural Network Toolbox
Artificial Neural [22, 23]. As shown in Fig. 1, there are three layers in the network:
Network QSAR an input layer, a hidden layer, and an output layer. Three different
(FANN-QSAR) types of molecular fingerprints, namely FP2 [24], MACCS [25],
Input Layer Hidden Layer Output Layer

(MACCS, FP2 or ECFP6 Fingerprint)
1
.
0 .
. pKi or pIC50
.
. .
.
0
1
Bias Bias
Fig. 1 Graphical representation of the fingerprint-based ANN-QSAR (FANN-QSAR) model. (Reprinted with per-
mission from ref. 17. Copyright (2012) American Chemical Society)
and Extended-Connectivity Fingerprint (ECFP6) [26], were used

in this study. The generation of fingerprints is described in
Subheading 3. The number of input layer neurons was equal to the
size of the fingerprint. For example, FP2 and ECFP6 fingerprints
have 1,024 bits and therefore, the number of input neurons is
equal to 1,024. Similarly, there are 256 input neurons for the
MACCS fingerprint. The number of hidden layer neurons was var-
ied between 100 and 1,000. The networks were trained using gra-
dient descent with momentum training function (traingdm) to
update weights and biases, the tangent sigmoid transfer function
(tansig) for the hidden layer and the linear transfer function (pure-
lin) for the output layer. 10 % randomly selected compounds from
the training data were used as a validation set to decide when to
stop training. The model training was stopped after 4,000 epochs
(iterations) or if the mean square error (MSE) of prediction on the
training set had reached the minimum value of 0.1. In addition,
early stopping was enabled when the prediction error on the vali-
dation set increased for 300 epochs and the weights and biases at
the minimum of the validation error were returned. The optimal
number of hidden neurons was selected via cross-validation experi-
ments in which the model was trained using different numbers of
hidden neurons, and an average of training set and validation set
mean squared errors (MSE) was calculated. The number of hidden
neurons which gave the lowest average MSE was used as the opti-
mal number for subsequent model testing on the test set. The
mean squared error (MSE) is defined as
1 T
∑ ( t (i ) − p (i ) )
2
MSE =
T i =1
where T is the total number of training samples, t(i) is the target

value of the ith sample, and p(i) is the predicted value of the ith
sample.
2.3 Comparison The performance of the FANN-QSAR was compared to those of

of the FANN-QSAR other reported 3D- and 2D-QSAR methods, including CoMFA
Model with Other [27], CoMSIA [28], Hologram QSAR (HQSAR) [29, 30], QSAR
Methods by eigenvalue analysis (EVA) [31], back-propagation feed-forward
neural network implemented in Cerius2 using 2.5D descriptors
(NN 2.5D), and ensemble neural network [32] (NN-ens) using
2.5D descriptors which were implemented and tested by Sutherland
et al. [21]. Three different fingerprints, namely FP2, ECFP6, and
MACCS, were used as inputs for FANN-QSAR models, and each
model was trained separately for each fingerprint type. During each
training process, a cross-validation experiment was performed to
decide the optimal number of hidden neurons which was used sub-
sequently on the test set prediction. To compare objectively, the
FANN-QSAR models were trained and tested on the same training
and test data sets provided by Sutherland et al. [21]. Three FANN-
QSAR models were named as follows based on which molecular
fingerprint was used as an input: ECFP6-ANN-QSAR, FP2-ANN-
QSAR, and MACCS-ANN-QSAR. Results of CoMFA, CoMSIA
basic, HQSAR, EVA, NN (2.5D), and NN-ens (2.5D) methods
were taken from the work of Sutherland et al. [21].
Final correlation coefficient (r2 test) values of each data set are
listed in Table 3. Comparisons of r2 (test) values across all data sets
show that ECFP6 fingerprint-based ANN-QSAR model (ECFP6-
ANN-QSAR) performed better than FP2 and MACCS fingerprint-
based models for all data sets. For ACE, AchE, and COX2 data
sets, the CoMFA model performed better than ECFP6-ANN-
QSAR model but by a small margin. The ECFP6-ANN-QSAR
model performed better for the DHFR and BZR data sets.
Performance of the CoMSIA model was similar to that of the
ECFP6-ANN-QSAR model. It is important to note that CoMFA
Table 3
Comparison of results from different QSAR methods
ECFP6- FP2-ANN- MACCS- CoMSIA NN NN-ens

ANN-QSAR QSAR ANN-QSAR CoMFA basic HQSAR EVA (2.5D) (2.5D)
ACE 0.41 0.2 0.08 0.49 0.52 0.3 0.36 0.39 0.51
AchE 0.43 0.13 0.04 0.47 0.44 0.37 0.28 −0.04 0.21
BZR 0.31 0.08 0.06 0 0.08 0.17 0.16 0.39 0.34
COX2 0.28 0.22 0.23 0.29 0.03 0.27 0.17 0.31 0.32
DHFR 0.63 0.43 0.48 0.59 0.52 0.63 0.57 0.42 0.54
and CoMSIA are field-based 3D QSAR methods which require

similar scaffolds and high-quality molecular alignments to make
effective predictions [11]. On the other hand, ECFP6-ANN-
QSAR is a fingerprint-based method which works on structurally
diverse data sets and requires no alignment during the model train-
ing process, which makes it more robust and high throughput in
virtual screening. However, different fingerprints can produce dif-
ferent results and, in our work, ECFP6 produced an overall better
result across different data sets compared to FP2 and MACCS fin-
gerprints. In addition to 3D QSAR methods, the FANN-QSAR
models were compared to another 2D QSAR method known as
hologram QSAR (HQSAR), which is based on molecular holograms
containing counts of molecular fragments similar to fingerprints.
It can be observed that ECFP6-ANN-QSAR performed consis-
tently better than HQSAR in all data sets except for DHFR data
set resulting in the same r2 test value (0.63). The FANN-QSAR
models were also compared to other neural network approaches
that used 2.5D descriptors as reported by Sutherland et al. The
ECFP6-ANN-QSAR model performed better than the NN (2.5D)
method in three out of five data sets and an ensemble of ten neural
networks (NN-ens) approach using 2.5D descriptors performed
slightly better than ECFP6-ANN-QSAR model in three out of five
data sets. It is important to note that all QSAR models failed for
COX2 and BZR data sets (r2 test < 0.34) and had moderate perfor-
mances (r2 test < 0.64) for the other three data sets. Overall, the
ECFP6-ANN-QSAR model performed consistently across all data
sets and its performance was comparable to other 3D, 2D, and
neural networks QSAR methods previously reported.
2.4 Cannabinoid We extended the application of FANN-QSAR by predicting the

Receptor Binding binding activities of cannabinoid ligands. A total of 1,699 struc-
Activity Prediction turally diverse cannabinoid ligands with reported CB2 binding
for Known affinities were used. The CB2 binding activity data were down-
Cannabinoid Ligands loaded from the CBID data set compiled by the Xie lab (http://
www.cbligand.org/cbid/index.php). The ligands were randomly
divided into training and test sets. FANN-QSAR models using dif-
ferent fingerprints were trained on training sets and the optimal
numbers of hidden neurons were selected via cross-validation.
Figure 2 contains a summary of cross-validation results for all three
FANN-QSAR models. It can be observed that different training
and test sets as well as different types of fingerprints resulted in dif-
ferent optimal numbers of hidden neurons, which suggested that
cross-validation experiments are necessary to train neural networks
for the best results.
After such training and parameter tuning, the predictive accu-
racy of the final model on the test set was evaluated. The process
was repeated five times and a summary of r2 values from each round
of experiments can be seen in Table 4. Within each round, the
0.8
Round 1 ECFP6- Optimal No. of Training Validation Average
Round 2 ANN-QSAR Hidden Neurons Set MSE Set MSE MSE
0.7 Round 3
Round 4 Round 1 1000 0.182 0.795 0.488
Average MSE
0.6 Round 5
Round 2 600 0.257 0.749 0.503
0.5 Round 3 200 0.176 0.593 0.384
0.4 Round 4 900 0.223 0.594 0.409
Round 5 300 0.151 0.638 0.394

0.3
100 200 300 400 500 600 700 800 900 1000
Round 1
0.9
FP2-ANN- Optimal No. of Training Validation Average
Round 2
QSAR Hidden Neurons Set MSE Set MSE MSE
Round 3
0.8 Round 1 700 0.300 0.809 0.554
Round 4
Round 5
Average MSE
0.7 800 0.360 0.713 0.537

Round 2
0.6
Round 3 800 0.360 0.466 0.413
0.5
Round 4 700 0.305 0.716 0.511
0.4
Round 5 700 0.290 0.606 0.448
0.3
100 200 300 400 500 600 700 800 900 1000
Round 1
Round 2
0.9 MACCS- Optimal No. of Training Validation Average
Round 3
Round 4 ANN-QSAR Hidden Neurons Set MSE Set MSE MSE
0.8 Round 5 Round 1 800 0.354 0.781 0.567
0.7 Round 2 600 0.384 0.783 0.583

Average MSE
0.6 Round 3 900 0.361 0.617 0.489
Round 4 800 0.352 0.749 0.550

0.5
Round 5 800 0.339 0.605 0.472
0.4
100 200 300 400 500 600 700 800 900 1000
Num. of Hidden Neurons
Fig. 2 Cross-validation results of each FANN-QSAR method on CB2 ligand data set. (Reprinted with permission
from ref. 17. Copyright (2012) American Chemical Society)
same training and test compounds were used across all three
FANN-QSAR models. For example, the same training and test
compounds in Round 1 of the ECFP6-ANN-QSAR model were
used in the Round 1 of the FP2-ANN-QSAR and MACCS-ANN-
QSAR models. As shown in the table, the ECFP6-ANN-QSAR
model consistently outperformed the FP2- and MACCS-ANN-
QSAR models in all five rounds of experiments. The ECFP6-
ANN- QSAR model achieved an average r2 test value of 0.56
(r = 0.75) across all repeat experiments compared with 0.48 (r = 0.69)
and 0.45 (r = 0.67) for the FP2- and MACCS-ANN-QSAR mod-
els, respectively. Results showed that the ECFP6 fingerprint was
better than FP2 and MACCS fingerprints for the cannabinoid data
set as well as the other five data sets. In fact, it has been also
reported that circular fingerprints such as ECFP6 fingerprints are
found to be more useful in virtual screening and ADMET properties’
Table 4
A summary of the performance of each FANN-QSAR model on CB2 ligand
data set
Round r2 Training r2 Test

ECFP6-ANN-QSAR
1 0.86 0.55
2 0.81 0.63
3 0.87 0.53
4 0.84 0.56
5 0.89 0.54
FP2-ANN-QSAR
1 0.78 0.55
2 0.74 0.60
3 0.74 0.38
4 0.77 0.46
5 0.79 0.40
MACCS-ANN-QSAR
1 0.74 0.48
2 0.72 0.53
3 0.74 0.37
4 0.74 0.47
5 0.75 0.41
prediction studies [33, 34]. Our results suggested that an ECFP6

fingerprint-based ANN-QSAR model can be used in virtual screen-
ing of chemical ligands in a high-throughput manner since it only
requires 2D fingerprints as inputs instead of 3D molecular align-
ments and bioactive conformations, as required by other 3D QSAR
methods.
2.5 Cannabinoid To more rigorously test the predictive ability of the FANN-QSAR
Receptor Binding method on new cannabinoid compounds which are not in the Xie
Activity Prediction group’s cannabinoid ligand training data set, the most recently
on Newly Reported reported cannabinoid ligands and associated CB2 binding affinity
Cannabinoid Ligands data were downloaded from ChEMBL database [35]. These com-
pounds were not found in the training (CBID) data set and were
collected to be used as a new test set in order to evaluate the
FANN-QSAR performance. The new test data set consisted of
295 compounds with reported CB2 Ki values which were then
12
11 rtest = 0.75
10
9
Predicted pKi
8
7
6
5
4
3
3 4 5 6 7 8 9 10 11 12
Experimental pKi
Fig. 3 Scatter plot between experimental pKi and predicted pKi values of 278 test cannabinoid ligands after the
removal of 17 outliers. (Reprinted with permission from ref. 17. Copyright (2012) American Chemical Society)
converted to pKi values. 41.55 % of new CB2 ligands were less

than 80 % similar (2D Tanimoto similarity) and 25.34 % were less
than 70 % similar to the training compounds. This similarity analy-
sis indicated that the newly reported CB2 compounds contained a
good mixture of similar and dissimilar compounds to the training
database. The ECFP6-ANN-QSAR model was trained using the
1,699 CB2 ligand (CBID) data set. Twenty independent rounds
of training and testing were performed. For each round, a ran-
domly selected 90 % of the database was used for training and the
remaining 10 % was used for validation. As a result, 20 indepen-
dent trained models were derived. After 20 rounds of predictions,
an average predicted value for each test compound was calculated.
The average residual value was 0.046 and the standard deviation
was 1.03. Seventeen outlier compounds with residuals more than
two standard deviations away from the average residual were
removed. Figure 3 shows a scatter plot of experimental and pre-
dicted pKi values of 278 test compounds after such outlier
removal. The linear regression of these 278 data points provided
an r of 0.75, slope of 0.686, and y intercept of 2.249. This result
indicated that there was a good correlation between experimental
and predicted values, given the fact that many of these test com-
pounds have novel structures and were not included in the model
training and validation process. The result suggested that the
FANN-QSAR possessed good generalization ability for newly
reported cannabinoid ligands.
2.6 Virtual Screening In order to illustrate how the FANN-QSAR model could be used
of the NCI Compound in drug discovery research, we applied it as a virtual screening
Database for Lead tool to search for CB2 lead ligands from the NCI compound data-
Cannabinoid Ligands base [36]. For consistency, the same 20 trained models in the
previous section were used. The NCI database, containing 329,089

compounds, was filtered to remove duplicate compounds, isotopes,
metals, and mixtures using the Tripos Selector program [30, 37].
This filtering reduced it to 211,782 compounds that were used as
a test set for each round of prediction. For each compound the
ECFP6 fingerprint was generated and used as the network input
to predict the CB2 receptor binding activity. After 20 rounds of
predictions, an average predicted value for each compound was
calculated. The top ranked 50 compounds were selected, but
only 10 compounds were physically available from the NCI via
material transfer agreement (MTA). These ten compounds were
experimentally tested for CB2 activities using a [3H]CP-55940
competition binding assay experiment. The experimental proto-
col for this validation assay is described in Subheading 3.
Among the ten tested NCI compounds, four (NSC49888,
NSC174122, NSC369049, and NSC76301) had CB2 Ki between
6.70 nM (pKi = 8.17) and 3.80 μM (pKi = 5.42). One compound,
which has a similar chemical scaffold to the well-known cannabi-
noid ligand, delta-9-tetrahydrocannabinol, was found to be a high-
affinity compound with an average CB2 Ki value of 6.70 nM
(pKi = 8.17). These four compounds and other similar compounds
(70 % 2D Tanimoto similarity threshold was used) [38] were not
found in the training database. Among the top 50 ligands, there
was one NCI compound (NSC768843) which was more than
90 % similar (Tanimoto coefficient ≥ 0.9) to a known classical can-
nabinoid ligand (CAS ID: 112830-95-2 or HU210), an analog of
delta-9-tetrahydrocannabinol, reported in the literature [39].
These findings proved that the FANN-QSAR method can find not
only novel compounds with good CB2 binding affinities but also
compounds similar to known ligands from a testing database con-
taining thousands of compounds with diverse scaffolds. Hit ligands
with novel scaffolds can be used as lead compounds for further
medicinal chemistry optimization and SAR studies, while hits simi-
lar to known ligands provide additional information for scaffold
hopping and R-group variations which may be useful for medicinal
chemists. Table 5 contains the structures of NCI hit compounds
and their experimental pKi as well as predicted values. Apart for
one compound (NSC746843) that was not available from NCI,
the other four compounds were experimentally tested in our lab
and competition binding curves are shown in Fig. 4.
It should be noted that the predicted pKi correlated well with
experimental pKi for two of the five hit ligands but not for the
other three ligands. This finding could be attributed to the experi-
mental variability of the reported CB2 binding activities of train-
ing compounds among different research labs, or to a possible
limitation of 2D fingerprint descriptors which considers individual
Table 5
Identified NCI hit compounds with CB2 binding activities
Structure NSC ID MW ClogP Experimental pKi Predicted pKi

CH3
H3C
CH3
746843 400.55 6.61 8.81a 8.66
O
CH3
CH3
CH O
HO
CH3
49888 330.46 5.59 8.17b 8.28
O O
H3C
CH3
CH3 CH3
OH
O
174122 463.52 4.76 5.59c 8.41
N
N O
H3C CH3
O
H3C
N
H3C
369049 488.66 4.00 5.51c 8.48
O
O H3C H3C
CH3
O
H3 C H3C
O HO
H3C
O
76301 354.44 3.99 5.42c 8.21
a
An average literature reported Ki value of a known cannabinoid compound (HU210) which is more than 90 % similar
to 746843
b
An average Ki value of two independent experiments performed in duplicate
c
An experimental Ki value of one experiment performed in duplicate
Fig. 4 CB2 receptor binding affinity Ki values of four NCI hit compounds measured by [3H]CP-55940 radioli-
gand competition binding assay using human CB2 receptors harvested from transfected CHO-CB2 cells.
(Reprinted with permission from ref. 17. Copyright (2012) American Chemical Society)
fragment contributions but sometimes may not be as effective as

other 3D descriptors when considering the overall structure of a
ligand. Fingerprints such as ECFP6 have, however, been found to
be useful in this study as well as in other several cheminformatics
studies [5, 33, 34], and they are known to be robust and time
efficient for high-throughput virtual screening applications where
hundreds of thousands of chemicals are involved, as in this study.
To conclude, results from the virtual screening exercise that was
validated experimentally demonstrated that the derived FANN-
QSAR model is capable of successfully identifying lead CB2 com-
pounds with good binding affinities as well as compounds similar
to known cannabinoid ligands, and providing additional insights
for R-group and scaffold hopping of known ligands.
3 Notes
We used three different types of molecular fingerprints, namely

FP2 [24], MACCS [25], and Extended-Connectivity Fingerprint
(ECFP6) [26]. FP2 is a path-based fingerprint which indexes
molecular fragments and MACCS is a key-based fingerprint which
uses 166 predefined keys, whereas ECFP6 is a circular topological
fingerprint which is derived using a variant of the Morgan algo-
rithm [40]. FP2 and MACCS fingerprints were generated using
the “babel” command from the OpenBabel program [24] while
ECFP6 fingerprints were generated using the “generatemd” com-
mand from the ChemAxon program (http://www.chemaxon.
com). Ligand chemical structures stored in SDF format were
used as inputs to generate fingerprints. For each ligand, polar
hydrogens were added using the OpenBabel program [24] before
fingerprint generation. All fingerprints were fixed-length binary
representations with 1,024 bits for both ECFP6 and FP2, and
256 bits for MACCS fingerprint. Fingerprints were generated for
each ligand in the data sets and used as inputs to train the FANN-
QSAR models.
In order to evaluate CB2 binding activity of virtually screened
ligands, competition binding assays were performed by displacing
radioactive [3H]CP-55940 radioligand. The experimental proto-
col has been established based on previously reported procedures
[41–44] and is described briefly below.
A Perkin Elmer 96-well TopCounter is used in our laboratory
to measure the CB receptor binding affinity (Ki) of the in silico-
screened ligands by displacing [3H]CP-55940. In competition
binding experiments, ligands were diluted in dilution buffer
(50 mM Tris, 5 mM MgCl2, 2.5 mM EGTA) containing 0.1 %
(w/v) fatty acid-free bovine serum albumin (BSA), 10 % dimethyl
sulfoxide, and 0.4 % methyl cellulose. Various concentrations of
ligands/samples are added in the same volume to 2.5 nM [3H]
CP-55940. Incubation buffer (50 mM Tris, 2.5 mM EGTA,
5 mM MgCl2, 0.1 % (w/v) fatty acid-free BSA) and cell mem-
brane preparations from CHO cells expressing CB2 receptors
(5 μg per well) are added to a final volume of 200 μL. For the
saturation binding experiments, varying concentrations of [3H]
CP-55940 (0.05–4 nM) with or without 5 μM of an unlabeled
known ligand (CP-55940) were incubated with the receptor
membrane preparations to determine Kd and nonspecific binding.
After the binding suspensions are incubated at 30 °C for 1 h, the
reaction is terminated by rapid filtration through microfiltration
plates (Unifilter GF/B filterplate, Perkin Elmer) followed by five
washes with ice-cold TME buffer containing 0.1 % BSA on a
Packard Filtermate Harvester (Perkin Elmer). The plates are then
dried overnight and 30 μl MicroScint 0 scintillation liquid is added
to each well of the dried filter plates. Then the bound radioactivity
is counted using a Perkin Elmer 96-well TopCounter. The Ki is
calculated by using nonlinear regression analysis (Prism 5;
GraphPad Software Inc., La Jolla, CA), with the Kd values for
[3H]CP-55940 determined from saturation binding experiments.
This assay is used for determining binding affinity parameters (Ki)
of ligand-receptor interactions between the CB2 receptor and
ligands.
References
1. Myint KZ, Xie X-Q (2010) Recent advances in pharmacophoric approach. SAR QSAR Environ
fragment-based QSAR and multi-dimensional Res 22(5–6):525–544. doi:10.1080/1062936x.
QSAR methods. Int J Mol Sci 11(10): 2011.569948
3846–3866 11. Chen J-Z, Han X-W, Liu Q et al (2006)
2. Perkins R, Fang H, Tong W et al (2003) 3D-QSAR studies of arylpyrazole antagonists of
Quantitative structure-activity relationship cannabinoid receptor subtypes CB1 and CB2.
methods: perspectives on drug discovery and A combined NMR and CoMFA approach. J Med
toxicology. Environ Toxicol Chem 22(8): Chem 49(2):625–636
1666–1679 12. Vilar S, Santana L, Uriarte E (2006)
3. Salum L, Andricopulo A (2009) Fragment- Probabilistic neural network model for the in
based QSAR: perspectives in drug design. Mol silico evaluation of anti-HIV activity and
Divers 13(3):277 mechanism of action. J Med Chem 49(3):
4. Chen JZ, Wang J, Xie XQ (2007) GPCR 1118–1124. doi:10.1021/jm050932j
structure-based virtual screening approach for 13. González-Díaz H, Bonet I, Terán C et al (2007)
CB2 antagonist search. J Chem Inf Model ANN-QSAR model for selection of anticancer
47(4):1626–1637. doi:10.1021/ci7000814 leads from structurally heterogeneous series of
5. Wang L, Ma C, Wipf P et al (2013) compounds. Eur J Med Chem 42(5):580–585.
TargetHunter: an in silico target identification doi:10.1016/j.ejmech.2006.11.016
tool for predicting therapeutic potential of 14. Patra JC, Chua BH (2011) Artificial neural net-
small organic molecules based on chemoge- work-based drug design for diabetes mellitus
nomic database. AAPS J 15:395–406, PMID: using flavonoids. J Comput Chem 32(4):
23292636 555–567. doi:10.1002/jcc.21641
6. Tandon M, Wang L, Xu Q et al (2012) A tar- 15. Dimitrov I, Naneva L, Bangov I et al (2014)
geted library screen reveals a new inhibitor Allergenicity prediction by artificial neural net-
scaffold for protein kinase D. PLoS One 7(9): works. J Chemometrics. doi:10.1002/cem.2597
e44653. doi:10.1371/journal.pone.0044653, 16. Vanyúr R, Héberger K, Kövesdi I et al (2002)
PMID: 23028574 Prediction of tumoricidal activity and accumu-
7. Ma C, Wang L, Yang P et al (2013) LiCABEDS lation of photosensitizers in photodynamic
II. Modeling of ligand selectivity for G-protein therapy using multiple linear regression and
coupled cannabinoid receptors. J Chem Inf artificial neural networks. Photochem
Model 53(1):11–26. doi:10.1021/ci3003914 Photobiol 75(5):471–478. doi:10.1562/0031-
8. Wang L, Ma C, Wipf P et al (2012) Linear and 8655(2002)0750471potaaa2.0.co2
nonlinear support vector machine for the clas- 17. Myint K-Z, Wang L, Tong Q et al (2012)
sification of human 5-HT1A ligand functional- Molecular fingerprint-based artificial neural
ity. Mol Inf 31(1):85–95. doi:10.1002/ networks QSAR for ligand biological activity
minf.201100126 predictions. Mol Pharm 9(10):2912–2923.
9. Myint K, Xie X-Q (2011) Fragment-based doi:10.1021/mp300237z
QSAR algorithm development for compound 18. Molnár L, Keserű GM (2002) A neural network
bioactivity prediction. SAR QSAR Environ Res based virtual screening of cytochrome P450
22(3):385–410 3A4 inhibitors. Bioorg Med Chem Lett
10. Chen JZ, Myint KZ, Xie X-Q (2011) New 12(3):419–421. doi:10.1016/s0960-894x(01)
QSAR prediction models derived from GPCR 00771-5
CB2-antagonistic triaryl bis-sulfone analogues 19. Muresan S, Sadowski J (2005) “In-House
by a combined molecular morphological and Likeness”: comparison of large compound
collections using artificial neural networks. 34. Glem R, Bender A, Arnby C et al (2006)
J Chem Inf Model 45(4):888–893. doi: Circular fingerprints: flexible molecular descrip-
10.1021/ci049702o tors with applications from physical chemistry
20. Wang L, Xie XQ (2012) Cannabinoid Ligand to ADME. IDrugs 9(3):199–204
Database. Accessed Nov 2011 35. Bellis LJ, Akhtar R, Al-Lazikani B et al (2011)
21. Sutherland JJ, O’Brien LA, Weaver DF (2004) Collation and data-mining of literature bioactiv-
A comparison of methods for modeling quanti- ity data for drug discovery. Biochem Soc Trans
tative structure–activity relationships. J Med 39(5):1365–1370. doi:10.1042/BST0391365,
Chem 47(22):5541–5554. doi:10.1021/ BST0391365 [pii]
jm0497141 36. Collins J, Crowell J (2011) Drug Synthesis and
22. Greenidge PA, Carlsson B, Bladh L-G et al Chemistry Branch, Developmental Therapeutics
(1998) Pharmacophores incorporating numer- Program (DTP), Division of Cancer Treatment
ous excluded volumes defined by x-ray crystallo- and Diagnosis, National Cancer Institute.
graphic structure in three-dimensional database http://dtp.nci.nih.gov/
searching: application to the thyroid hormone 37. Tripos (2012) SYBYL-X 1.2. 1699 South
receptor. J Med Chem 41(14):2503–2512 Hanley Rd., St. Louis, Missouri, 63144, USA
23. Mathworks (2007) MATLAB. 7.5.0.342 38. Xie XQ, Chen JZ (2008) Data mining a small
(R2007b) edn, Natick, MA molecule drug screening representative subset
24. O’Boyle N, Banck M, James C et al (2011) from NIH PubChem. J Chem Inf Model
Open Babel: an open chemical toolbox. 48(3):465–475. doi:10.1021/ci700193u
J Cheminform 3:33 39. Huffman JW, Yu S, Showalter V et al (1996)
25. Durant JL, Leland BA, Henry DR et al (2002) Synthesis and pharmacology of a very potent
Reoptimization of MDL keys for use in drug cannabinoid lacking a phenolic hydroxyl with
discovery. J Chem Inf Comput Sci 42(6):1273– high affinity for the CB2 receptor. J Med Chem
1280. doi:10.1021/ci010132r 39(20):3875–3877. doi:10.1021/jm960394y
26. Rogers D, Hahn M (2010) Extended- 40. Morgan HL (1965) The generation of a unique
connectivity fingerprints. J Chem Inf Model machine description for chemical structures—a
50(5):742–754. doi:10.1021/ci100050t technique developed at Chemical Abstracts
27. Cramer R, Patterson D, Bunce J (1988) Service. J Chem Doc 5(2):107–113. doi:
Comparative molecular field analysis (CoMFA). 10.1021/c160017a018
1. Effect of shape on binding of steroids to car- 41. Gertsch J, Leonti M, Raduner S et al (2008)
rier proteins. J Am Chem Soc 110:5959–5967 Beta-caryophyllene is a dietary cannabinoid.
28. Klebe G (1998) Comparative molecular simi- Proc Natl Acad Sci 105(26):9099–9104.
larity indices analysis: CoMSIA. 3D QSAR in doi:10.1073/pnas.0803601105
Drug Design. Three-Dimensional Quantitative 42. Raduner S, Majewska A, Chen J-Z et al (2006)
Structure Activity Relationships 3:87–104 Alkylamides from Echinacea are a new class
29. Lowis D (1997) HQSAR: a new, highly predic- of cannabinomimetics: cannabinoid type 2
tive QSAR technique. Tripos Technical Notes receptor-dependent and -independent immu-
1(5):17 nomodulatory effects. J Biol Chem 281(20):
14192–14206
30. Jain AN (2004) Ligand-based structural
hypotheses for virtual screening. J Med Chem 43. Zhang Y, Xie Z, Wang L et al (2011) Mutagenesis
47(4):947–961 and computer modeling studies of a GPCR con-
served residue W5.43(194) in ligand recogni-
31. Ferguson AM, Heritage T, Jonathon P et al tion and signal transduction for CB2
(1997) EVA: a new theoretically based molecu- receptor. Int Immunopharmacol 11(9):1303–
lar descriptor for use in QSAR/QSPR analysis. 1310. doi:10.1016/j.intimp.2011.04.013
J Comput Aided Mol Des 11(2):143–152. doi
:10.1023/a:1008026308790 44. Yang P, Wang L, Feng R et al (2013) Novel
triaryl sulfonamide derivatives as selective can-
32. Agrafiotis DK, Cedeño W, Lobanov VS (2002) nabinoid receptor 2 inverse agonists and osteo-
On the use of neural network ensembles in clast inhibitors: discovery, optimization, and
QSAR and QSPR. J Chem Inf Comput Sci biological evaluation. J Med Chem
42(4):903–911. doi:10.1021/ci0203702 56(5):2045–2058. doi:10.1021/jm3017464
33. Bender A, Jenkins JL, Scheiber J et al (2009) 45. DePriest SA, Mayer D, Naylor CB et al (1993)
How similar are similarity searching methods? 3D-QSAR of angiotensin-converting enzyme
A principal component analysis of molecular and thermolysin inhibitors: a comparison of
descriptor space. J Chem Inf Model 49(1):108– CoMFA models based on deduced and experi-
119. doi:10.1021/ci800249s mentally determined active site geometries.
J Am Chem Soc 115(13):5372–5384. doi: inflammatory agents. J Med Chem 38(22):

10.1021/ja00066a004 4570–4578. doi:10.1021/jm00022a023
46. Sugimoto H, Tsuchiya Y, Sugumi H et al 55. Reitz DB, Li JJ, Norton MB et al (1994)
(1992) Synthesis and structure-activity rela- Selective cyclooxygenase inhibitors: novel
tionships of acetylcholinesterase inhibitors: 1,2-diarylcyclopentenes are potent and orally
1-benzyl-4-(2-phthalimidoethyl)piperidine, active COX-2 inhibitors. J Med Chem 37(23):
and related derivatives. J Med Chem 35(24): 3878–3881. doi:10.1021/jm00049a005
4542–4548. doi:10.1021/jm00102a005 56. Khanna IK, Yu Y, Huff RM et al (2000) Selective
47. Sugimoto H, Tsuchiya Y, Sugumi H et al cyclooxygenase-2 inhibitors: heteroaryl modi-
(1990) Novel piperidine derivatives. Synthesis fied 1,2-diarylimidazoles are potent, orally
and anti-acetylcholinesterase activity of 1-benzyl- active antiinflammatory agents. J Med Chem
4-[2-(N-benzoylamino)ethyl]piperidine deriv- 43(16):3168–3185. doi:10.1021/jm0000719
atives. J Med Chem 33(7):1880–1887. doi: 57. Khanna IK, Weier RM, Yu Y et al (1997)
10.1021/jm00169a008 1,2-diarylimidazoles as potent, cyclooxygen-
48. Haefely W, Kyburz E, Gerecke M et al (1985) ase-2 selective, and orally active antiinflamma-
Recent advances in the molecular pharmacol- tory agents. J Med Chem 40(11):1634–1647.
ogy of benzodiazepine receptors and in the doi:10.1021/jm9700225
structure-activity relationships of their agonists 58. Khanna IK, Weier RM, Yu Y et al (1997)
and antagonists. Adv Drug Res 14:165–322 1,2-Diarylpyrroles as potent and selective
49. Chavatte P, Yous S, Marot C et al (2001) inhibitors of cyclooxygenase-2. J Med Chem
Three-dimensional quantitative structure– 40(11):1619–1633. doi:10.1021/jm970036a
activity relationships of cyclo-oxygenase-2 59. Gangjee A, Vidwans AP, Vasudevan A et al
(COX-2) inhibitors: a comparative molecular (1998) Structure-based design and synthesis of
field analysis. J Med Chem 44(20):3223–3230. lipophilic 2,4-diamino-6-substituted quinazo-
doi:10.1021/jm0101343 lines and their evaluation as inhibitors of dihy-
50. Talley JJ, Brown DL, Carter JS et al (2000) drofolate reductases and potential antitumor
4-[5-Methyl-3-phenylisoxazol-4-yl]- agents. J Med Chem 41(18):3426–3434.
benzenesulfonamide, Valdecoxib: a potent and doi:10.1021/jm980081y
selective inhibitor of COX-2. J Med Chem 60. Rosowsky A, Mota CE, Wright JE et al (1994)
43(5):775–777. doi:10.1021/jm990577v 2,4-Diamino-5-chloroquinazoline analogs of
51. Huang H-C, Li JJ, Garland DJ et al (1996) trimetrexate and piritrexim: synthesis and anti-
Diarylspiro[2.4]heptenes as orally active, folate activity. J Med Chem 37(26):4522–
highly selective cyclooxygenase-2 inhibitors: 4528. doi:10.1021/jm00052a011
synthesis and structure–activity relationships. 61. Rosowsky A, Cody V, Galitsky N et al (1999)
J Med Chem 39(1):253–266. doi:10.1021/ Structure-based design of selective inhibitors
jm950664x of dihydrofolate reductase: synthesis and anti-
52. Penning TD, Talley JJ, Bertenshaw SR et al parasitic activity of 2,4-diaminopteridine ana-
(1997) Synthesis and biological evaluation of logues with a bridged diarylamine side chain.
the 1,5-diarylpyrazole class of cyclooxygenase- J Med Chem 42(23):4853–4860. doi:10.1021/
2 inhibitors: identification of 4-[5-(4-
jm990331q
Methylphenyl)-3-(trifluoromethyl)-1H- 62. Graffner-Nordberg M, Kolmodin K, Åqvist J
pyrazol-1-yl]benzenesulfonamide (SC-58635, et al (2001) Design, synthesis, computational
Celecoxib). J Med Chem 40(9):1347–1365. prediction, and biological evaluation of ester
doi:10.1021/jm960803q soft drugs as inhibitors of dihydrofolate reduc-
53. Li JJ, Norton MB, Reinhard EJ et al (1996) tase from Pneumocystis carinii. J Med Chem
Novel terphenyls as selective cyclooxygenase-2 44(15):2391–2402. doi:10.1021/jm010856u
inhibitors and orally active anti-inflammatory 63. Gangjee A, Elzein E, Queener SF et al (1998)
agents. J Med Chem 39(9):1846–1856. Synthesis and biological activities of tricyclic
doi:10.1021/jm950878e conformationally restricted tetrahydropyrido
54. Li JJ, Anderson GD, Burton EG et al (1995) annulated furo[2,3-d]pyrimidines as inhibitors
1,2-Diarylcyclopentenes as selective cyclooxy- of dihydrofolate reductases. J Med Chem
genase-2 inhibitors and orally active anti- 41(9):1409–1416. doi:10.1021/jm9705420
Chapter 10
GENN: A GEneral Neural Network for Learning Tabulated

Data with Examples from Protein Structure Prediction
Eshel Faraggi and Andrzej Kloczkowski
Abstract
We present a GEneral Neural Network (GENN) for learning trends from existing data and making
predictions of unknown information. The main novelty of GENN is in its generality, simplicity of use, and
its specific handling of windowed input/output. Its main strength is its efficient handling of the input data,
enabling learning from large datasets. GENN is built on a two-layered neural network and has the option
to use separate inputs–output pairs or window-based data using data structures to efficiently represent
input–output pairs. The program was tested on predicting the accessible surface area of globular proteins,
scoring proteins according to similarity to native, predicting protein disorder, and has performed remark-
ably well. In this paper we describe the program and its use. Specifically, we give as an example the constru-
ction of a similarity to native protein scoring function that was constructed using GENN. The source code
and Linux executables for GENN are available from Research and Information Systems at http://mamiris.
com and from the Battelle Center for Mathematical Medicine at http://mathmed.org. Bugs and problems
with the GENN program should be reported to EF.
Key words Neural network, Protein scoring, Windowed input, Automatic learning, GENN, Protein
structure prediction
1 Introduction
Proteins perform their function through structure-based spatial

and temporal interactions and give rise to the biosphere as we
know it. More knowledge about them will contribute to our fun-
damental understanding of life and in practical terms will revolu-
tionize medicine, bioscience, and other fields. However, in the
context of present day computational and analytical tools they are
too large and too complex. This is best exemplified in the ever
growing gap between the number of known protein structures
(relatively few) and the number of known protein sequences (rela-
tively many). As of the end of 2013, there are under 100,000
structures deposited in the Protein Data Bank but over 33 million
165
166 Eshel Faraggi and Andrzej Kloczkowski
protein sequences from about 30,000 different organisms in the

NCBI RefSeq database. In many other areas of human endeavor
constantly growing amount of information is being collected.
Frequently, this information can be useful in other, related or unre-
lated, human endeavors [1–15]. Consumer purchasing history is
an example where the past purchases of a consumer group can be
used to assign a probability distribution of future purchases, and
could, for example, give advertisers critical information for market-
ing strategy. In protein structure prediction too, experimentally
solved protein structures can be used to infer the structure of
unsolved proteins.
These vast amounts of data require specialized tools to extract
useful information from them. Here we present one such tool that
can be used to analyze large amounts of information and learn
from it based on examples. The learning is achieved through the
steepest descent training of a two layer neural network. It is able to
efficiently handle instance-based and window-based training as will
be described below. Its novelty is in its ability to handle any quan-
titative information, limited only by hardware, and that it can
efficiently store window-based data, enabling modification of the
modeling approaches without the need to modify the database.
For large window-based training these are unique features as far
as we could find. We term this tool GENN for GEneral Neural
Network.
GENN was programmed in FORTRAN 90. In its design and
implementation its efficiency and usability were of major concern.
It is constructed out of several subroutines that process the data,
initialize the model, and train it. It is also capable of producing
predictions from existing single models or producing ensemble
predictions with expected deviations. Its execution is terminal
based. It was built on Ubuntu Linux, under the BASH (Bourne
Again SHell) environment. For the rest we will use that environ-
ment to describe usage.
Although GENN can handle any data, it was designed and built
for questions related to the relationship between the residue
sequence of a protein and its native structure. Therefore we shall use
terminology from that field to describe various features of the pro-
gram. In general terms the problem is as follows. We are given the
residue sequence of a protein, which can be derived from the genetic
code. This residue sequence is in essence the chemical formula of
the protein molecule. In turn we are to provide the best prediction
for the physical structure of this molecule. However, since this is a
difficult problem, certain approximations are usually made, either
by coarse graining the full atom model, or by looking at derived
structural properties such as the accessible surface area [16–26].
The general architecture of GENN is given in Fig. 1. The user
supplied information includes the feature files which include
information that will be used to establish a prediction, and the
GEneral Neural Network 167
Fig. 1 Schematic diagram describing the architecture of GENN. The feature files
include information used to establish a prediction. The training files contain infor-
mation to be predicted. The error between the neural network predicted value
and the training values is used to train its weights. H1 and H2 refer to the first
and second hidden layers, respectively
training files which contain the information that is to be predicted.

The error between the neural network predicted value and the
training values is used to train the weights using the steepest
descent back propagation method [21]. These weights are between
the inputs through hidden layers one and two (H1 and H2, respec-
tively, in Fig. 1) and to the prediction output.
2 Types of Training and Prediction in GENN
2.1 Instance-Based We use the term instance-based prediction to describe cases in which
Prediction each example to be learned from is independent of every other
example in the database. This is in contrast to windows-based pre-
diction, to be discussed next, where examples covered in overlapping
windows will share common features. All programs related to
instance-based prediction have a base name of “genn2inst.”
An example of instance-based prediction is prediction of global
protein features. When one wants to predict quantities for entire
protein sequences, for example the radius of gyration, one is deal-
ing with instance-based prediction. For each protein, various input
features are gathered but in general different proteins will have
unique input features that are not shared by other proteins.
Another example we have implemented using GENN, that will be
described below, is that of predicting the fitness of model protein
structures to the native state. Such a scheme was implemented suc-
cessfully using GENN for the CASP10 experiment [27].
2.2 Window-Based Window-based prediction is used to describe cases where different

Prediction instances, with different outputs, share common inputs. Maybe the
easiest case to describe is that of predictions for properties of resi-
dues along a protein sequence. In this case, we use sequentially
neighboring residues (within a window of a certain width) to
obtain the input features, then as we go along the chain neighbor-
ing residues will share common input features. In this case it is
wasteful to record the inputs for each instance (residue) separately.
It is more efficient to only store the input/output features of the
protein and extract the appropriate inputs and outputs appropriate
for each residue. All programs related to window-based prediction
have a base name of “genn2wind.” More discussion of window-
based prediction is given in earlier works [21, 25, 28–31].
Boundary conditions are important for window-based predic-
tion. For example, for residues near a protein terminus the sliding
window may fall out of range. These considerations are internally
controlled in GENN. Windows that encompass regions outside
the boundary of the instance are truncated to include only those
parts that fall within range.
Because of the relative complexity associated with partitioning
a database of windowed instances we designed the genn2wind ver-
sion to automatically perform an n-fold cross validation, where n is
determined by the command line options. By default n is equal to
ten and the last partition is taken as the test set. In this mode 5 %
out of the remaining training data is used for over-fit protection.
If one wishes to use only an over-fit protection set without a test
set, for example while developing a server, this can be achieved
using the “-sv” option. This built-in n-fold cross validation mode
is limited to the window-based programs. For the instance-based
programs only an over-fit set is chosen by default.
3 Database and Initialization File
In general the database is composed of separate directories each

representing a given instance, a protein in our example. Each direc-
tory contains the input and output files associated with each
instance. In our example the output file contains a list of the nor-
malized ASA values to be predicted for each residue. The exact
details of this normalization will be given later in the text when
RASA will be defined. The input files then should have quantities
that have predictive power with respect to the desired output.
Specific details of the input and output files will be given when
discussing ASAquick.
The initialization file is used to describe the locations of the
required information for a given run. An example is provided in
Fig. 2. The first line in the file gives the path of the head directory
Fig. 2 Example of the input file for GENN. The first line gives the head directory where all the instance directo-
ries are. This should be enclosed in double quotes to preserve the slashes. The next line describes the inputs
used separated by spaces. The third line lists the outputs to be predicted represented by file names separated
by spaces. The remaining lines list the instances on which the training will be performed
where all the instance directories are located. This should be

enclosed in double quotes to preserve the slashes in the path name.
The next line describes the inputs used, which are assumed the
same for all instances. Each input is represented by a file name that
will contain its values for a particular instance in its instance direc-
tory. Multiple inputs should be separated by spaces. In a similar
way the third line lists the outputs to be predicted, represented by
files in the instance directory that contains these values. Multiple
outputs should be separated by spaces. The design of GENN is
intended to facilitate changing both inputs and outputs for testing
and research. The remaining lines list the instances on which the
training will be performed. For the case of instance-based predic-
tion this lists individual instances. For window-based prediction
this lists a collection of instances grouped together in single files,
for example the residues in a given protein. To allow for user
control GENN does not modify the order of the list of instances.
Randomization of the order of training instances, which can be
useful, is left to the user.
One should note that for window-based prediction, since each
input file can contain many instances (residues) an additional
requirement on the database is imposed. To ascertain a match
between inputs and outputs for a given instance two identifiers are
used (originally the crystallographer index and the residue type for
proteins). During the reading of the database both identifiers are
checked and the program halts on mismatches between different
files.
4 Training and Prediction
The specifics of using GENN for training and predicting from data
are given below.
4.1 Training a Model The program is run from the command line. The options associ-
ated with training are given in Table 1. By default the program
looks for a file called “genn.in” to read the training dataset structure
from (initialization file). If this file does not exist in the directory it
is necessary to include in the command line a “-l” option followed
by the name of the initialization file.
Given additional data a model can be retrained by including it
as the initial condition for the new round of training. This is
achieved using the “-wf” option followed by the name of the file
containing the weights and parameters of the trained model.
In this case the weights and other model parameters are used as
initial values and no randomization is carried out. Then new train-
ing is performed on the new instances.
4.2 Predictions Once a model has been trained all parameters are stored in a file.
from Trained Models This file then can be used to give a prediction for new instances.
In addition a collection of models can be used to produce an
average prediction with an accompanying standard deviation. This
standard deviation can prove beneficial in cases where quality of
prediction is desired, but this requires testing on specific cases. The
basic options associated with prediction from a trained model are
given in Table 1. Note that some options are common between
training and prediction tasks, however, model parameters such as
the activation parameter or others that are related to the model
architecture are stored in, and read from, the model file and should
be reissued only in the uncommon event that one desires to over-
ride their values.
5 Special Options
Several unique features are provided with GENN. These options

grew out of several applications related to protein structure predic-
tion [21, 25, 29–31]. In this section we outline their possible uses.
Refer to Table 1 for the required syntax. These options are available
for the window-based programs.
5.1 Filter Network Some cases for window-based prediction can benefit from a filter,
where predictions over a window are used as inputs for a follow-up
prediction. Examples include protein secondary structure where a
filter prediction layer is often used [21, 25, 29]. While a filter layer
can always be run using a second-stage neural network, for the
Table 1
GENN options
Categorya Flagb Description Defaultc

-l db list file genn.in
-owf Weights output file
-m Maximum number of epochs 1000
-np Number of database files to use whole db
-r Random seed time dependent
-cv Test fold index 10
-f Fraction of database for test fold 0.1
-wi Input window size 21
-wo Output window size 1
-h1 Number of hidden layer one nodes 21
-h2 Number of hidden layer two nodes 21
Train -mi Maximum number of bad iterations 100
-a Activation parameter 0.2
-u Learning rate 0.001
-p Momentum 0.4
-hf Number of filter hidden layer nodes 11
-nf Run network filter
-gf Guiding factor 2.d0
-ng Don’t use distance guiding
-gi Number of global inputs 0
-go Number of global outputs 0
-df Use degeneracy files
-pr1 Predict single file (give id)
Predict -prl File of ids to predict
-dw Not same weight types, reread database
-aw Average over weights in file
-d Database head directory
-wf Weights input file
Both -so Average over outputs (nvo)
-po Probability output
-h Print help
Options available for GENN
a
We distinguish three types of categories for options. “Train” for training specific options. “Predict” for prediction
specific options. And, “universal” for options that used in both
b
The flag or command line option
c
Default values, some flags either do not have a default but require a value or are logical and do not require a value
window-based programs a filter layer can be included by using an

option in the command line. Refer to Table 1 for the appropriate
syntax. By default the filter network runs with 11 nodes, this value
can be changed as needed. Note that a specific name is used to save
the filter model. When getting predictions from trained filter mod-
els the filter option should also be used.
5.2 Guided Learning In some instances of window-based prediction the relevance of a

given location to a particular prediction site can be dependent on
the distance between the location and the prediction site. In these
cases it may be helpful to guide the network in that direction. One
way of achieving this is by introducing an extra set of weights that
have such distance dependence [21]. This is the default behavior
for the window-based program. An option is available to switch off
this behavior.
5.3 Global Inputs For certain window-based prediction global variables may be
and Outputs important. We shall consider the case of protein secondary struc-
ture to illustrate the point. In a window-based approach the sec-
ondary structure of a given residue is considered in the context
of a window around it. However, certain parameters such as the
amino acid composition of the protein may contain information
about general tendencies to form a specific protein class (all-alpha,
all-beta, alpha/beta, or alpha + beta). This information would typ-
ically not be contained in a limited window view and hence can
help the prediction. Indeed global input features can significantly
contribute to the prediction accuracy. To use global inputs/outputs,
one should store their values in files called genn.gin/genn.gut,
respectively, in the directory corresponding to that instance.
An option, “-gi” for global input and/or “-go” for global output, is
required in the command line to include these global files. These
options should be followed by the number of values to use from
these global files enabling control of how much global information
to use. Note that options for global files are only required for train-
ing. The input structure is recorded in the weight files and is
retrieved from them when called to predict.
5.4 Degeneracy In some cases of window-based prediction one may like to sample
Training certain cases more often than others. For example, for disordered
proteins one may like to sample more of the disordered residues in a
given protein instance [31]. In instance-based prediction one can
include instances and their occurrence at will by repeating them in the
file “genn.in.” For window-based prediction to change the sampling
frequency within an instance, one includes a file with an integer count
of the number of times (including zero) to train on a given training
output. Hence, the degeneracy file order should match the rest of the
input/output and should include the corresponding indexes.
Degeneracy files should be stored in instance directories under files
named “genn.dgn,” and an option should be given for their use.
5.5 Types of Output GENN is able to predict multiple outputs. This in turn enables
specific functionality besides general multi-output prediction. The
first option is to treat the multiple outputs as components of
a multi-state probability vector and train, optimize, and report
accordingly. While training is not affected by this option, optimiza-
tion is carried on the ability of the trained network to correctly
distinguish the most dominant component of the probability
vector. Reporting also follows the same protocol. This behavior
was found useful when an ancestor of GENN was used to predict
protein secondary structure [25].
Another possibility is to predict several instances of the same
output and generate an average prediction from the same trained
network. This has proven beneficial in several instances of continuous
value prediction [29]. It provides an extra layer of averaging, with
more control over variation between prediction models and hence
the ability to average over specific random errors in certain circum-
stances. This also generates an estimate for the prediction stability in
the reported standard deviation over predictions and in this form may
be beneficial in assigning prediction stability and accuracy.
6 Automated Learning
GENN was started as the first, and simplest, component of an auto-

mated learning machine. Automated learning here means being
able to “sit” on a big database (e.g., the Internet) and answer in
meaningful ways to questions posed by either humans or machines.
A sort of Watson [wikipedia.org/wiki/Watson_(computer)] but
without human design. Rather, continuously, progressively, and
automatically developing an image of the database pertinent to the
questions it was exposed to in its existence. Here we specifically
mean a non-memorizing learner, it is expected that on certain ques-
tions, different learners will respond differently depending on their
histories. In this respect GENN was set up to create local networks
on the fly.
The other, more complex component of such a learning
machine would require to extract information in a meaningful way
to describe both input and output features of a general question/
answer pair. Note that some questions have multiple acceptable
answers. Also, it would require building a self-sustaining learning
mechanism. Both these tasks are monumental and would require
breaking down the various problems further.
7 Examples
In this section we describe work that grew out of this version of

GENN. Older versions of GENN helped produce an NMR fluc-
tuation predictor [30] and a protein disorder predictor termed
SPINE-D [31], as well as some other applications. GENN was also

involved in various testing in several labs. SPINE-D participated in
the CASP9 experiment and was ranked among the top five meth-
ods [32]. This version of GENN was also used to create ASAquick,
a fast accessible surface area predictor that uses only single sequence
information [33], and a knowledge-based protein scoring function
termed Seder [27] which we will present here as an example of use
of GENN. Seder participated in the CASP10 [32] experiment as
the “Kloczkowski_Lab” group. It was ranked second in predicting
the structure of the “hard” targets category [34], third for “all”
targets. It was the only prediction approach that was ranked in the
top three for both “hard” and “all” cases.
This is a good place to point out that for any practical problem,
the machine learner is only secondary to the problem of represen-
tation and interpretation of the input and the output. Care and
thought of how to present the input and extract the output
can produce great improvements in learning and prediction quality.
To a second degree, given the same input/output information,
different quality learners can produce different outcomes. For
example, it has been our general experience that smaller sets of
data are better learnt by methods such as support vector machines
while larger sets are better learnt by neural networks, the difference
being several relative percent accuracy. On the other hand,
for example, changes of the representation of the input/output for
predicting dihedral angles of proteins has resulted in almost 100 %
reduction in the prediction error [18, 21, 28, 29].
Seder [27] is an example of an implementation of GENN.
It was developed specifically for structured proteins, to rank struc-
tural models according to their similarity to a native structure.
In the directory “example/Seder” of the source code distribution
we give examples for the feature files used to make Seder. We use
the same names to introduce them here, more information is avail-
able in the Seder paper [27]. The first line of the input file “genn.
in” gives the directory location of the database of instances. In the
second line of “genn.in” the inputs are listed. typ233w.norm con-
tains information about the distance of individual atoms to the
solvent. There are a total of 936 real numbers making up this
feature. res2res refers to the partial energy sums between pairs of
residues. There are a total of 441 real numbers making up this
feature. 4bod, dfire2, and rwplus refer to the four-body [35–37],
DFIRE2 [38, 39], and RWPlus [40] potentials, respectively. The
desired output from the neural network, given on the third line, is
a transformation of the TM-score [41, 42] used to measure simi-
larity between native and model structures (decoys). For training
we used server models from the Critical Assessment of protein
Structure Prediction (CASP) [32] rounds 5 through 9 (94,717
structures). The fourth line of “genn.in” points to the single
example of an instance given here. This points to subdirectories of
the “db” directory. Actual training will involve a list of such

instances. We average over several training realizations with differ-
ent initial conditions. For each, in addition to randomizing the
initial weights, we also randomize the list of training proteins (PDB
+ CASP) and select thirty thousand of them. From this subset,
30 % were used to compose the over-fit-protection set while the
remaining 70 % was used for online training. Over-fit testing was
done after each training epoch.
The command line for training one realization of a Seder neu-
ral network is given by “nohup time./genn2inst.e -l seder.in -f 0.3
-mi 200 -h1 51 -h2 31 -r $RANDOM &.” We have included three
features of Linux/BASH that are not directly related to this work
but are nonetheless useful. The first is the “nohup” command
which allows a training of the weights even after logout or loss of
connection. Note that you must have the ampersand symbol at the
end of the line for this functionality. The second is the “time” com-
mand which is useful in estimating run times and optimization.
The third is “$RANDOM” which is a BASH reserved word for
generation of pseudorandom numbers, it comes in handy in auto-
mation of neural network creation since seeding is done indepen-
dently of your application and depends on the system state (hence
arbitrary). The training command line given above will train a set
of weights and output them with other necessary information for
prediction into a file. By design the weights file name is constructed
uniquely from the process ID and other parameters associated with
the run. Here we shall assume its name to be “wei.out.” A single
prediction from this file for a single protein with ID, PID, would
look something like “./genn2inst.e -wf wei.out -pr1 PID.” Where
is it assumed that the directory mentioned in the beginning of wei.
out contains a subdirectory called PID with the necessary files.
Note that GENN is designed to calculate prediction error, hence it
will look for target files even if asked to predict. If such a target file
does not exist, is not in your possession, or a complete blind test of
the software is sought, trivial (zero filled) target files should be
generated. Of course then the reported prediction errors are mean-
ingless. For a collection of proteins with PIDs listed in the file “list.
PID” and taking the average over a collection of weight files listed
in file “list.wei” prediction and prediction variation are obtained
via the command “./genn2inst.e -aw list.wei -prl list.PID.” Both
these prediction methods produce a text output to screen which
can be piped to a text file. If a list of proteins is given the first col-
umn in the output is the PID listed in “list.wei.” The program
“awk” can then be easily used to generate individual prediction
files if those are necessary. Note that we have given an example of
instance-based prediction, and the syntax for window-based pre-
diction is identical. Example files for window-based prediction are
in the subdirectory “db/pdb.window.”
8 Summary
We have presented GENN, a general neural network designed

to train on ad-hoc data. GENN was designed with efficiency and
modularity in mind as part of a more complex algorithm of an
automated learner. It can take any numerical input/output prob-
lem and prepare a corresponding, non-memorizing, model struc-
ture to represent this data. Data can be organized in files containing
individual instances or a collection of ordered instances where each
line is an individual input/output target. GENN is available from
Research and Information Systems at http://mamiris.com, and
from the Battelle Center for Mathematical Medicine at http://
mathmed.org.
Acknowledgements
We gratefully acknowledge the financial support provided by

the National Institutes of Health (NIH) through Grants
R01GM072014 and R01GM073095 and the National Science
Foundation through Grant NSF MCB 1071785. Both authors
would like to thank the organizers of CASP10 conference in Gaeta,
Italy, for inviting them to the conference and providing free regis-
tration to EF. EF would also like to thank Yaoqi Zhou and Keith
Dunker for hosting him at IUPUI and general discussions.
References
1. Kassin SM (1979) Consensus information, resources variables: a review of modelling issues
prediction, and causal attribution: a review of and applications. Environ Model Softw 15:
the literature and issues. J Pers Soc Psychol 101–124
37:1966 8. Chang JC, Wooten EC, Tsimelzon A,
2. Crick NR, Dodge KA (1994) A review and Hilsenbeck SG, Gutierrez MC, Elledge R,
reformulation of social information-processing Mohsin S, Osborne CK, Chamness GC, Allred
mechanisms in children’s social adjustment. DC et al (2003) Gene expression profiling
Psychol Bull 115:74 for the prediction of therapeutic response to
3. Fielding AH, Bell JF (1997) A review of meth- docetaxel in patients with breast cancer. Lancet
ods for the assessment of prediction errors in 362:362–369
conservation presence/absence models. Environ 9. Schofield W et al (1985) Predicting basal
Conserv 24:38–49 metabolic rate, new standards and review of
4. Makhoul J (1975) Linear prediction: a tutorial previous work. Hum Nutr Clin Nutr 39:5
review. Proc IEEE 63:561–580 10. Blundell T, Sibanda B, Sternberg M, Thornton
5. Fontenot RJ, Wilson EJ (1997) Relational J (1987) Knowledge-based prediction of pro-
exchange: a review of selected models for a pre- tein structures. Nature 326:26
diction matrix of relationship activities. J Bus 11. Chou PY, Fasman GD (1978) Empirical pre-
Res 39:5–12 dictions of protein conformation. Annu Rev
6. Rost B et al (2001) Review: protein secondary Biochem 47:251–276
structure prediction continues to rise. J Struct 12. Floudas C, Fung H, McAllister S, Mönnigmann
Biol 134:204–218 M, Rajgaria R (2006) Advances in protein
7. Maier HR, Dandy GC (2000) Neural networks structure prediction and de novo protein design:
for the prediction and forecasting of water a review. Chem Eng Sci 61:966–988
13. Moult J (2005) A decade of CASP: progress, 26. Wang C, Xi L, Li S, Liu H, Yao X (2012)
bottlenecks and prognosis in protein structure A sequence-based computational model for the
prediction. Curr Opin Struct Biol 15:285–289 prediction of the solvent accessible surface area
14. Vazquez A, Flammini A, Maritan A, Vespignani for α-helix and β-barrel transmembrane resi-
A (2003) Global protein function prediction dues. J Comput Chem 33:11–17
from protein-protein interaction networks. Nat 27. Faraggi E, Kloczkowski A (2013) A global
Biotechnol 21:697–700 machine learning based scoring function for
15. Borgwardt KM, Ong CS, Schönauer S, protein structure prediction. Proteins: Struct
Vishwanathan S, Smola AJ, Kriegel H-P (2005) Funct Bioinf. doi:10.1002/prot.24454
Protein function prediction via graph kernels. 28. Xue B, Dor O, Faraggi E, Zhou Y (2008) Real
Bioinformatics 21:i47–i56 value prediction of backbone torsion angles.
16. Chothia C (1974) Hydrophobic bonding and Proteins: Struct Funct Bioinf 72:427–433
accessible surface area in proteins. Nature 29. Faraggi E, Yang Y, Zhang S, Zhou Y (2009)
248:338–339 Predicting continuous local structure and the
17. Moret M, Zebende G (2007) Amino acid effect of its substitution for secondary structure
hydrophobicity and accessible surface area. in fragment-free protein structure prediction.
Phys Rev E 75:011920 Structure 17:1515–1527
18. Dor O, Zhou Y (2007) Real-spine: an inte- 30. Zhang T, Faraggi E, Zhou Y (2010) Fluct-
grated system of neural networks for real-value uations of backbone torsion angles obtained
prediction of protein structural properties. from nmr-determined structures and their
Proteins: Struct Funct Bioinf 68:76–81 prediction. Proteins: Struct Funct Bioinf 78:
19. Durham E, Dorr B, Woetzel N, Staritzbichler 3353–3362
R, Meiler J (2009) Solvent accessible surface 31. Zhang T, Faraggi E, Xue B, Dunker AK,
area approximations for rapid and accurate pro- Uversky VN, Zhou Y (2012) Spine-d: accurate
tein structure prediction. J Mol Model 15: prediction of short and long disordered regions
1093–1108 by a single neural-network based method.
20. Zhang H, Zhang T, Chen K, Shen S, Ruan J, J Biomol Struct Dyn 29:799–813
Kurgan L (2009) On the relation between resi- 32. Moult J, Fidelis K, Kryshtafovych A,
due flexibility and local solvent accessibility in Tramontano A (2011) Critical assessment
proteins. Proteins: Struct Funct Bioinf 76: of methods of protein structure prediction
617–636 (casp) round ix. Proteins: Struct Funct Bioinf
21. Faraggi E, Xue B, Zhou Y (2009) Improving 79:1–5
the prediction accuracy of residue solvent 33. Faraggi E, Yaoqi Z, Kloczkowski A (2014)
accessibility and real-value backbone torsion Accurate single-sequence prediction of solvent
angles of proteins by guided-learning through accessible surface area using local and global
a two-layer neural network. Proteins: Struct features. Proteins: Struct, Funct, and Bioinf.
Funct Bioinf 74:847–856 DOI: 10.1002/prot.24682
22. Zhang T, Zhang H, Chen K, Ruan J, Shen S, 34. CASP10 (2012) Official group performance
Kurgan L (2010) Analysis and prediction of ranking. http://www.predictioncenter.org/
RNA-binding residues using sequence, evolu- casp10/groups_analysis.cgi. Accessed 10 June
tionary conservation, and predicted secondary 2012
structure and solvent accessibility. Curr Protein 35. Feng Y, Kloczkowski A, Jernigan R (2007)
Pept Sci 11:609–628 Four-body contact potentials derived from two
23. Gao J, Zhang T, Zhang H, Shen S, Ruan J, protein datasets to discriminate native struc-
Kurgan L (2010) Accurate prediction of pro- tures from decoys. Proteins: Struct Funct
tein folding rates from sequence and sequence- Bioinf 68:57–66
derived residue flexibility and solvent 36. Feng Y, Kloczkowski A, Jernigan RL (2010)
accessibility. Proteins: Struct Funct Bioinf Potentials’ r’us web-server for protein energy
78:2114–2130 estimations with coarse-grained knowledge-
24. Nunez S, Venhorst J, Kruse CG (2010) based potentials. BMC Bioinf 11:92
Assessment of a novel scoring method based 37. Gniewek P, Leelananda SP, Kolinski A, Jernigan
on solvent accessible surface area descriptors. RL, Kloczkowski A (2011) Multibody
J Chem Inf Model 50:480–486 coarse-grained potentials for native structure
25. Faraggi E, Zhang T, Yang Y, Kurgan L, Zhou recognition and quality assessment of protein
Y (2012) Spine x: improving protein secondary models. Proteins: Struct Funct Bioinf
structure prediction by multistep learning 79:1923–1929
coupled with prediction of solvent accessible 38. Zhou H, Zhou Y (2002) Distance-scaled, finite
surface area and backbone torsion angles. ideal-gas reference state improves structure-
J Comput Chem 33:259–267 derived potentials of mean force for structure
selection and stability prediction. Protein Sci selection and structure prediction. PLoS One
11:2714–2726 5:e15386
39. Yang Y, Zhou Y (2008) Ab initio folding of 41. Zhang Y, Skolnick J (2004) Scoring function
terminal segments with secondary structures for automated assessment of protein structure
reveals the fine difference between two closely template quality. Proteins: Struct Funct Bioinf
related all-atom statistical energy functions. 57:702–710
Protein Sci 17:1212–1219 42. Xu J, Zhang Y (2010) How significant is a pro-
40. Zhang J, Zhang Y (2010) A novel side-chain tein structure similarity with tm-score = 0.5?
orientation dependent potential derived from Bioinformatics 26:889–895
random-walk reference state for protein fold
Chapter 11
Modulation of Grasping Force in Prosthetic Hands

Using Neural Network-Based Predictive Control
Cristian F. Pasluosta and Alan W.L. Chiu
Abstract
This chapter describes the implementation of a neural network-based predictive control system for driving
a prosthetic hand. Nonlinearities associated with the electromechanical aspects of prosthetic devices pres-
ent great challenges for precise control of this type of device. Model-based controllers may overcome this
issue. Moreover, given the complexity of these kinds of electromechanical systems, neural network-based
modeling arises as a good fit for modeling the fingers’ dynamics. The results of simulations mimicking
potential situations encountered during activities of daily living demonstrate the feasibility of this
technique.
Key words Nonlinear control, Prosthetic hands, Neural networks, Optimization
1 Introduction
A prosthetic hand must be able to exercise precise control of many

degrees of freedom (DoF) to maintain stable grasping for various
scenarios in the activities of daily living (ADL). A prosthetic hand
must be also designed to avoid psychological and muscle fatigue
(human–machine interaction design) and must meet the electro-
mechanical specifications such as anthropomorphic dimensions,
weight, and power consumption (mechatronic design).
The device control approach can be classified in four directions
(Fig. 1). One approach is to control the dynamics of the hand
using direct information decoded from bio-signals, with vision as
the only feedback modality (Fig. 1a) [1–9]. In an attempt to reduce
errors in control, a second approach has been proposed in which
artificial sensory information (i.e., tactile and proprioceptive) is
sent back to the user (Fig. 1b) [10–15]. A third approach has been
designed to reduce psychological effort and hence increase usabil-
ity and controllability of the device. In this case, the control struc-
ture is organized in a hierarchical fashion by dividing it into low
and high levels of control (Fig. 1c) [16–29]. In the high level of
179
180 Cristian F. Pasluosta and Alan W.L. Chiu
Fig. 1 Control approaches, modified from ref. 34, with permission
control, bio-signals are decoded to interpret the intention of the

user (for example, open/close the hand, finger pre-shape), and in
the low level of control a stand-alone feedback controller is used to
modulate the finger dynamics automatically. This approach feeds
artificial sensory information back to the controller but not to the
user. Finally, in an effort to combine the advantages of the second
and third approaches into a fourth control structure, researchers
have adopted the abovementioned hierarchical control strategy
with sensory feedback to the user (Fig. 1d) [30–33]. This chapter
focuses on the third approach. More specifically, we discuss the
implementation of a neural networks-based control system for the
low level of control.
Several prototypes with a hierarchical control structure have
been developed. Cipriani et al. developed a control strategy for the
CyberHand [23] that makes use of the information from an exten-
sive array of sensors placed at the prosthetic hand. This control
structure is divided into two stages. The first one consists of a pre-
shaping of the hand, whereas a position control system arranges
the fingers in a specific configuration (such as precision pinch, tri-
pod, etc.). In the second stage, a force control system adjusts the
grasping force around the target object. Another hierarchically
organized control strategy was presented by Light et al. and was
applied to control the Southampton-REMEDI hand [25]. In a
multi-stage fashion, the fingers first pre-shape and close around the
object while applying minimal force. When indicated by the user,
the system modulates the force applied by the fingers to produce
stable grasping and to avoid slippage. The object is then released
when the hand opens after a squeeze signal coming from the user.
Engeberg et al. have used the Motion Control hand to test several
control approaches, in which the user controls the grip force from
EMG signals directly, but at the same time taking advantage of a
force control system [19–21]. Although this is not a hierarchical
Neural Network Predictive Control 181
control structure per se, the modulation of the grip force using
artificial sensory information can be considered within this line of
research. In their approach, force-derivative feedback was intro-
duced into the force control loop to improve the sensibility and to
reduce the force overshoot at the beginning of the grasping [20].
A hybrid force-velocity-position control was also implemented to
address the overshoot issue [21]. Further, using the derivative of
the shear force, an adaptive slip prevention system was imple-
mented to improve the force control [19]. Another slip prevention
algorithm that also minimizes object deformation was presented
by Engeberg and Meek [27]. Andrecioli and Engeberg utilized an
adaptive controller based on the detection of object stiffness [28].
Wettels et al. implemented a control system that adjusts the grip
force according to a slippage detection system that uses a fluid-
based tactile sensor to compute the normal and tangential compo-
nents of the grip force [17]. Peerdeman et al. utilized the University
of Bologna hand IV to test a low level of control based on the
intrinsically passive controller technique [29].
Consider now the low-cost mechatronic design of a prosthetic
hand shown in Fig. 2 [35]. This is a five-fingered hand with an
opposable thumb, with each finger containing two phalanges and
their tips slightly bent. The thumb, index and middle fingers move
independently (active fingers), while the ring and little fingers
(passive fingers) are mechanically coupled to the middle finger.
The fingers are flexed by DC gear motors by means of a tendon-cable
Fig. 2 Mechatronic design of a low-cost prosthetic hand, modified from ref. 34,
with permission
system connected through a worm gear system attached to a pulley.

An extensor tendon cable attached to a spring extends the fingers
when the flexor cable is released. Force sensor resistors (FSR)
placed at the tips and resistive flex sensors (RFS) placed at the
dorsal part of the fingers are used to obtain force and flexion
information, respectively.
Focusing on the low level of control of the hand (from a con-
trol system perspective), a prosthetic hand can be represented as a
plant, in which inputs control the actuators (i.e., current motor)
and the outputs are the kinematics and kinetics of the hand (i.e.,
contact force, joint angle, and torque). Thus, prostheses are highly
nonlinear devices mainly due to the existence of motor dead bands,
friction and large gear ratios [20], and nonlinear responses typi-
cally observed in low-cost sensors. In order to control the force/
position of the fingers the controller must be able to deal with
these nonlinearities. In this case, a linear controller (i.e., a PID
controller) will have difficulties stabilizing the system for all the
possible scenarios.
One way of dealing with nonlinear systems is by modeling
their dynamics. A nonlinear model of the relationship between
inputs (i.e., motor current) and the outputs (i.e., grip force) can be
used to predict future behaviors and adjust the input accordingly.
Under these conditions, nonlinear model predictive control
(NMPC) presents an attractive solution. This technique employs a
nonlinear model to determine future controllable inputs over a
defined time horizon [36]. Then, if we can predict future outputs,
we can use them to compute the optimal input to maintain the
desired force over this horizon [36].
Creating an acceptable model of the system can be very chal-
lenging, especially if little is known about its dynamics. A potential
solution arises from a technique commonly referred to as “black
box modeling”. In black box modeling, several inputs and outputs
are presented to an algorithm that maps their relationship in an
iterative fashion [36]. In this context, artificial neural networks are
commonly used as a black box modeling technique given their ability
to map any nonlinear function to a specific degree of accuracy [37].
Neural networks-based NMPC has been demonstrated to be a
powerful tool in a wide variety of different scenarios [38–44].
Figure 3 shows a block diagram of an implementation of a
neural network-based NMPC for controlling the finger dynamics
of a prosthetic hand. Here u, y, ŷ , θ, and r represent the motor
current/voltage, the current fingertip force, the predicted finger-
tip force, the finger flex angle, and the force reference, respectively.
The force reference can be determined for example by a slippage
detection system so that when slippage is detected, the force
reference is changed to counteract this phenomenon. In the fol-
lowing sections, we describe the main two blocks of the NMPC:
the system model and the optimization. These procedures are
based mainly on the general methodology presented in [36, 45].
Neural Network
Model q
Flex Sensor
y Force Sensor
g u y
Optimization Motor
Controller Hand
Fig. 3 Neural network-based control scheme (from ref. 35, with permission)
y(t–2)
y(t–1)
Feedforward
u(t–2) Neural
Network y(t)
u(t–1)
q(t–2)
q(t–1)
Fig. 4 NNARX structure (from ref. 34, with permission)
2 Neural Network-Based Modeling
Neural networks can be used as predictor machines by using an

autoregressive external input structure (NNARX) (Fig. 4; θ is the
angle measured by the RFS, which is a non-controllable variable).
A feedforward architecture is utilized with its inputs delayed in
time. The neural network is trained with time-delayed inputs and
current outputs. It maps the relationship between past inputs and
current outputs, hence is capable of predicting future outputs
from current inputs. Different training approaches, such as the
Levenberg–Marquardt method [4], can be applied.
Using a hyperbolic tangent and linear function for the hidden
and output layer, respectively, the predictions of the NNARX
model of Fig. 4 are obtained as follows:
yˆ (t + i ) = lw  tanh ( iwp + b1 )  + b2 (1)

p =  y (t − 2 ) , y (t − 1) , u (t − 2 ) , u (t − 1) , θ (t − 2 ) , θ (t − 1)  (2)

Fig. 5 Experiment set up (from ref. 34, with permission)
where iw is the input-to-hidden weight matrix, lw is the

hidden-to- output weight matrix, p is the two-sample-delayed
input vector, and b1 and b2 are the bias terms
Designing the network structure includes finding the optimum
number of inputs (number of lags) and hidden layers. In general,
the rule of thumb is to produce a small testing error. However, it
is possible that with a high sample frequency (compared to the
dynamics of the system), a low test error may not necessarily lead
to a good model [36]. Therefore, in order for the dynamics of the
system to be modeled correctly, it is also required that no correla-
tion exists between the inputs and the training error and no auto-
correlation of the training error itself [36] (Fig. 5).
The following experiments were conducted to obtain training
and testing data in order to find the best model. We apply an input
signal to the motor to make the finger close around objects with
different shapes and compliances. A chirp signal is applied as the
input signal to the motor (following [36]):
u (t ) = u o + A sin (ωt t ) (3)

t
ωt = ωstart + (ωfinal − ωstart ) (4)
N
The input signal is swept with different values of uo and A as
well as with different values of ωstart and ωfinal to excite the full range
of variables involved in the model. The force and the angle are
measured for each active finger. The signal amplitudes of the train-
ing and testing data sets are finally scaled to zero mean and unity
variance. Using these data, several models are then created off-line
and the ones with the best performance (based on the abovemen-
tioned measurements) are chosen for the online implementation.
3 Optimization
In the optimization block an objective function (Eq. 5) is mini-

mized with respect to the future control inputs. This objective
function takes into account the error between the predictive out-
puts and the reference values, as well as the changes in the inputs
at each iteration.
N2 Nu
J (t , U (t ) ) = ∑ r (t + i ) − y (t + i ) + ρ ∑  ∆u (t + i − 1)  (5)
2 2
i =N 1 i =1

U (t ) = u (t )u (t + N u − 1) 
T
(6)

In Eqs. 5 and 6, the parameters N2, N1 and Nu are the predic-
tion horizon, minimum prediction horizon, and control horizon,
respectively. The constant ρ restricts the changes in the control
input [36]. The input u(t) is supplied to the motor at each sample
point and it is bounded for convergence purposes of the optimiza-
tion algorithm.
The algorithm presented in refs. 36, 45 can be used to mini-
mize the objective function (in other words, to find the optimum
u(t)). This algorithm utilizes a gradient descent technique where
the future set of control inputs is determined by the update rule
indicated in Eq. 7:
∂J
U ( t + 1) = U ( t ) − η ( t ) (7)
∂U (t )

α (r (t ) − y (t ) )
η ( t ) = η0 e (8)

The constant α is determined empirically. Expressing Eq. 5 in
matrix notation:
J (t , U (t ) ) = E (t ) E (t ) + ρ∆U (t ) ∆U (t )
T T
(9)

where:
E (t ) = e (t + 1) ,...,e (t + N 2 ) 
T
(10)

e (t + i ) = r (t + i ) − yˆ (t + i ) ; for i = 1, …, N 2 (11)

∆U (t ) =  ∆u (t ) , …, ∆u (t + N u − 1)
T
(12)

We have that

∂J ∂Y (t ) ∂∆U (t )
= 2E (t ) + 2 ρ∆U (t ) (13)
∂U (t ) ∂U (t ) ∂U (t )

  
Y (t ) =  y (t + 1) ,…, y (t + N 2 ) 
T
(14)

1 0 0 0
 −1 1 0 0 
∂∆U (t ) 
=0  N u × N u matrix (15)
∂U (t )  
 −1 1 0
 0 0 −1 1 

 ∂y (t + 1) 
 0 0 
 ∂u (t ) 
 ∂y (t + 2 )
∂y (t + 2 ) 
∂Y (t )  0 
=  ∂u (t ) ∂u (t + 1)  (16)
∂U (t )  
 
 ∂y (t + N 2 )
∂y (t + N 2 )

∂y (t + N 2 ) 
 
 ∂u (t ) ∂u (t + 1) ∂u (t + N u − 1) 

A recursive algorithm is presented by Noriega and Wang [45]
to calculate Eq. 16:
 
∂y (t + n ) ∂y (t + n − 1)  ∂fˆ ( p ) 
= 1 + , (17)
∂u (t + m − 1) ∂u (t + m − 1)  ∂yˆ (t + n − 1) 
 

∂y (t + n − 1) dp
= lw sech 2 ( iwp + b1 )  iw (18)
∂u (t + m − 1) du

∂fˆ ( p ) dp
= lw sec h 2 ( iwp + b1 )  iw  (19)
∂yˆ (t + n − 1) dy

Here, n = 1, …, N2, m = 1, …, N2 and,
dp
= [0, 0, 0, …, 1, 0, …,0]
T
(20)
du
dp
 = [1, 0, …, 0, 0, …,0]
T
(21)
dy
The motor power consumption and undesirable oscillations
are reduced by turning off the motor when the measured force
falls within a tolerance range (which we set at ±5 % of the refer-
ence value).
4 Testing the Neural Network-Based NMPC on a Single Prosthetic Finger
Once an optimum model of the finger dynamics has been obtained,

a simple way to evaluate the performance of the neural network-
based NMPC system is by testing its step response. In this proce-
dure, the finger is closed around different objects (Table 1), while a
Table 1
Objects used to evaluate the step response of the neural network-based
NMPC system
Object Weight (g) Diameter (cm)

Styrofoam cup 2 7
Small plastic bottle 15 6.5
Big plastic bottle 40 8.5
Aluminum cylinder 110 7.3
Adapted from ref. 34, with permission
force reference signal is applied to the control system. For the

results presented in this chapter, the nonlinear model of the dynam-
ics of the fingers was created using the Levenberg–Marquardt
method of the NNSYSID toolbox for Matlab [46].
Figure 6 shows a representative recording of the behavior of
the system as the reference signal is incremented in steps of differ-
ent heights, with an approximately time interval of 20 s (N = 5).
The performance metrics for each reference step are the average
closing time, the average overshoot, the average rise time (the time
in which the system reaches 85 % of the reference value), and the
average steady state error (SSE, absolute mean difference between
the finger force and the reference value during steady state). The
transient state, which precedes the steady state, is defined as the
time in which the peak value of the force oscillations is greater than
15 % of the reference value.
Once the finger closed around the object the rise time for each
step was around 0.5 s (Fig. 7) for most of (for the cylinder, at
0.89 N was 1.22 s). Although the response time of the finger is
significantly longer with the cylinder, the overall response time of
the control system is within the acceptable range (less than 1 s).
Maximum overshoot was found at contact and it was reduced
during the subsequent force steps (Fig. 8). Rigid objects (such as
the aluminum cylinder) gave rise to a smaller overshoot than softer
objects (such as the plastic bottle). In general, the overshoot
decreased as the applied force reference increased.
The experiments showed a small SSE in all the step sizes
(Fig. 9), with an incremental trend as the force reference increased.
This was expected as the motor was turned off when the measured
force was within 5 % of the reference force.
Overall, the force control system is capable of adjusting the
grasping force adequately. The response of the system is fast
enough to meet a rise time of approximately 500 ms, with an
overshoot small enough to avoid object damage. Regarding long
term force control, the SSE is small enough (~0.02–0.13 N) to
maintain the same force during an extended period of time.
Force reference and system output (Bottle trial 1)
1.5
Force (N)
0.5
0
0 10 20 30 40 50 60 70 80 90 100
time (sec)
Motor signal
2
Voltage
−2
−4
0 10 20 30 40 50 60 70 80 90 100
time (sec)
Fig. 6 Representative recording of the system step response (from ref. 34, with permission)
Fig. 7 Average closing and rise time for each object. Modified from ref. 34, with permission
Fig. 8 Average overshoot for each step and all objects. Modified from ref. 34, with permission
Fig. 9 Average steady state error for each object. Modified from ref. 34, with permission
5 Implementation on a Whole Hand Prototype
So far we have discussed how to control the force applied by an

artificial finger using a neural network-based NMPC. We describe
now how to incorporate this approach into a hierarchically struc-
tured control strategy of an artificial hand.
Fig. 10 Block diagram of the control strategy (modified from ref. 34, with permission)
Figure 10 shows implementation of the force control system

in the control of the hand prototype described early in this chapter.
This control strategy consists of five stages: Pre-shape, Closing, Force
Control, Detection, and an Open. The same scheme was applied to
drive the thumb, index and middle fingers independently.
The control strategy starts at the Pre-shape stage, where a
hand configuration (i.e., cylindrical, tip, lateral, etc.) is selected by
the decoding of the user’s bio-signals. This stage will determine
also which fingers will be involved in the grasping (i.e., in the tip
configuration only the thumb and index fingers are used).
Immediately after this stage, the system switches to the Closing
stage, where the fingers close around the object at maximum speed
until making contact with the object. Once contact with the object
is established, the Force Control stage comes into play to maintain
an initial minimum force over the object. In the Force Control
stage the neural network-based NMPC technique described earlier
is implemented to maintain this minimum force over the object.
The value of this minimal force is empirically determined to pre-
vent slippage and object deformation. This force will be maintained
while no unintended movement of the object is detected by the
Detection stage. If some undesired movement occurs, the refer-
ence force of the NMPC is increased in discrete steps until the
movement ceases. The Detection stage thus alternates with the
Force Control stage. The Detection stage is defined by the deriva-
tive of the normal force measured by the FSR placed at the tips of
the fingers (please refer to [35] for more detail).
Fig. 11 Whole hand experiments (modified from ref. 35, with permission)
The control strategy is tested in three different conditions that

simulate potential scenarios of daily life [35]. These conditions are
adopted from established testing methods [47–49]. A representa-
tive example for the first condition is shown in Fig. 11. The first
scenario involves applying both sudden and gradual changes to the
mass of the object while it is grasped with a transverse volar grip.
In the second scenario a sudden perturbation is made to the
grasped object while maintaining a pinch grip. Finally, the third
scenario tests the ability to adjust grasping forces while rotational
forces are applied to the grasped object. In all the three experi-
ments, the displacement of the object (i.e., the distance traveled by
the object after applying the disturbance) is recorded by a motion
sensor [35].
For the transverse volar grip experiments, the maximum average

displacement of 7.6 mm and the maximum average displacement
for the pulp pinch configuration of 3.05 mm are achieved. For the
torque experiment, the maximum average angle displacement of
10.7° at a load of 500 g (11 N cm) is obtained. A more detailed
discussion of these results can be found in ref. 35. Taken together,
these results suggest that the control strategy adjusted the grasping
force in response to the applied disturbance.
The performance assessment suggests that this low-cost pros-
thetic hand can effectively mimic and perform daily life tasks. The
neural-network based control system accounted for nonlinear
behaviors in a reliable fashion, leading to reproducible finger force
and fast response. Embedded in a hierarchically structured control
strategy, this property enabled object grasping with minimal slip-
page and deformation effects. We anticipate the use of neural-
network approaches could optimize the function of other prosthetic
devices where nonlinear behaviors compromise their performance.
6 Notes
The degree of object deformation and sliding rely heavily on the

relationship between the initial reference force and the force incre-
ment step. As a result, the choice of these two parameters is critical
in fine-tuning the performance of this control strategy. While this
methodology proved successful in the grasping of rigid objects, it
is anticipated that mimicking the grasping of softer objects would
require an on-the-fly adjustment of the initial reference force
applied so as to avoid a very high grasping force of the hand which
could lead to subsequent object deformation.
Acknowledgement
This work was supported by the NIH NCRR INBRE grant

P20RR016456, and the Louisiana Board of Regents RCS
LEQSF(2007-10)-RD-A-20.
References
1. Smith RJ, Tenore F, Huberdeau D et al (2008) hand based on principal components analysis.
Continuous decoding of finger position from Annual international conference of the IEEE
surface EMG signals for the control of pow- Engineering in Medicine and Biology Society,
ered prostheses. 30th Annual international 3–6 Sep 2009, pp 5022–5025
conference of the IEEE Engineering in 3. Jingdong Z, Li J, Shicai S et al (2006) A five-
Medicine and Biology Society, 20–25 Aug fingered underactuated prosthetic hand system.
2008, pp 197–200 Proceedings of the IEEE international confer-
2. Matrone G, Cipriani C, Secco EL et al (2009) ence on mechatronics and automation, 25–28
Bio-inspired controller for a dexterous prosthetic Jun 2006, pp 1453–1458
4. Jingdong Z, Zongwu X, Li J et al (2005) IEEE Trans Neural Syst Rehabil Eng 13(4):
Levenberg-Marquardt based neural network 468–472. doi:10.1109/TNSRE.2005.856072
control for a five-fingered prosthetic hand. 15. Lundborg G, Rosén B (2001) Sensory substi-
Proceedings of the IEEE international confer- tution in prosthetics. Hand Clin 17(3):
ence on robotics and automation, 18–22 Apr 481–488
2005, pp 4482–4487 16. Mangieri E, Ahmadi A, Maharatna K et al
5. Zajdlik J (2006) The preliminary design and (2008) A novel analogue circuit for control-
motion control of a five-fingered prosthetic ling prosthetic hands. IEEE biomedical circuits
hand. Proceedings of the international confer- and systems conference, 20–22 Nov 2008,
ence on intelligent engineering systems, 0-0 pp 81–84
0, pp 202–206
17. Wettels N, Parnandi AR, Moon JH et al (2009)
6. Dapeng Y, Jingdong Z, Yikun G et al (2009) Grip control using biomimetic tactile sensing
EMG pattern recognition and grasping force systems. IEEE/ASME Trans Mechatron 14(6):
estimation: improvement to the myocontrol of 718–723
multi-DOF prosthetic hands. IEEE/RSJ inter-
18. Tura A, Lamberti C, Davalli A et al (1998)
national conference on intelligent robots and
Experimental development of a sensory control
systems, 10–15 Oct 2009, pp 516–521
system for an upper limb myoelectric prosthesis
7. Tsujiuchi N, Takayuki K, Yoneda M (2004) with cosmetic covering. J Rehabil Res Dev
Manipulation of a robot by EMG signals using 35(1):14–26
linear multiple regression model. Proceedings
of the IEEE/RSJ international conference on 19. Engeberg ED, Meek SG (2008) Adaptive
intelligent robots and systems, vol 1992, 28 object slip prevention for prosthetic hands
Sep–2 Oct 2004, pp 1991–1996 through proportional-derivative shear force
feedback. IEEE/RSJ International conference
8. Jingdong Z, Zongwu X, Li J et al (2006) EMG on intelligent robots and systems, 22–26 Sep
control for a five-fingered underactuated pros- 2008, pp 1940–1945
thetic hand based on wavelet transform and
sample entropy. International conference on 20. Engeberg ED, Meek S (2008) Improved grasp
intelligent robots and systems, 9–15 Oct 2006, force sensitivity for prosthetic hands through
pp 3215–3220 force-derivative feedback. IEEE Trans Biomed
Eng 55(2):817–821
9. Weir RF, Ajiboye AB (2003) A multifunction
prosthesis controller based on fuzzy-logic tech- 21. Engeberg ED, Meek SG, Minor MA (2008)
niques. Proceedings of the 25th annual inter- Hybrid force-velocity sliding mode control of a
national conference of the IEEE Engineering prosthetic hand. IEEE Trans Biomed Eng
in Medicine and Biology Society, vol 1672, 55(5):1572–1581
17–21 Sep 2003, pp 1678–1681 22. Zhao DW, Jiang L, Huang H et al (2006)
10. Pylatiuk C, Mounier S, Kargov A et al (2004) Development of a multi-DOF anthropomor-
Progress in the development of a multifunc- phic prosthetic hand. International conference
tional hand prosthesis. 26th Annual interna- on robotics and biomimetics, 17–20 Dec 2006,
tional conference of the IEEE Engineering in pp 878–883
Medicine and Biology Society, 1–5 Sep 2004, 23. Cipriani C, Zaccone F, Stellin G et al (2006)
pp 4260–4263 Closed-loop controller for a bio-inspired
11. Cipriani C, Antfolk C, Balkenius C et al (2009) multi-fingered underactuated prosthesis.
A novel concept for a prosthetic hand with a Proceedings of the IEEE international confer-
bidirectional interface: a feasibility study. IEEE ence on robotics and automation, 15–19 May
Trans Biomed Eng 56(11):2739–2743. 2006, pp 2111–2116
doi:10.1109/TBME.2009.2031242 24. Carrozza MC, Cappiello G, Micera S et al
12. Panarese A, Edin BB, Vecchi F et al (2009) (2006) Design of a cybernetic hand for percep-
Humans can integrate force feedback to toes in tion and action. Biol Cybern 95(6):629–644
their sensorimotor control of a robotic hand. 25. Light CM, Chappell PH, Hudgins B et al
IEEE Trans Neural Syst Rehabil Eng 17(6): (2002) Intelligent multifunction myoelectric
560–567. doi:10.1109/TNSRE.2009.2021689 control of hand prostheses. J Med Eng Technol
13. Rodriguez-Cheu LE, Casals A (2006) Sensing 26(4):139–146
and control of a prosthetic hand with myoelectric 26. Connolly C (2008) Prosthetic hands from
feedback. The first IEEE/RAS-EMBS interna- Touch Bionics. Ind Rob 35(4):290–293
tional conference on biomedical robotics and 27. Engeberg ED, Meek SG (2013) Adaptive slid-
biomechatronics, 20–22 Feb 2006, pp 607–612 ing mode control for prosthetic hands to
14. Dhillon GS, Horch KW (2005) Direct neural simultaneously prevent slip and minimize
sensory feedback and control of a prosthetic arm. deformation of grasped objects. IEEE/ASME
Trans Mechatron 18(1):376–385. doi:10.1109/ process. Chinese Control Conference, 7–11

TMECH.2011.2179061 Aug 2006, pp 1143–1147
28. Andrecioli R, Engeberg ED (2013) Adaptive 40. Bandyopadhyay B (2005) Neural network
sliding manifold slope via grasped object stiff- based predictive controller (NNPC): an investi-
ness detection with a prosthetic hand. gation into its application in Textiles.
Mechatronics 23(8):1171–1179 International conference on computational
29. Peerdeman B, Fabrizi U, Palli G et al (2012) intelligence for modelling, control and auto-
Development of prosthesis grasp control sys- mation and international conference on intelli-
tems on a robotic testbed. 4th IEEE RAS & gent agents, web technologies and internet
EMBS international conference on biomedical commerce, 28–30 Nov 2005, pp 963–967
robotics and biomechatronics, 24–27 Jun 41. Jianbin H, Shaohua T, Vandewalle J (1993) One
2012, pp 1110–1115 step ahead predictive control of nonlinear sys-
30. Luo Z-z, Wang F, Wang R-c (2006) Study of tems by neural networks. Proceedings of interna-
multi-freedom myoelectric prostheses with tac- tional joint conference on neural networks, vol
tile sense. 27th Annual international conference 2763, 25–29 Oct 1993, pp 2761–2764
of the Engineering in Medicine and Biology 42. Soloway D, Haley P (2001) Aircraft reconfigu-
Society, 17–18 Jan 2006, pp 3004–3007 ration using neural generalized predictive con-
31. Cipriani C, Zaccone F, Micera S et al (2008) trol. Proceedings of the 2001 American Control
On the shared control of an EMG-controlled Conference, 2001, vol 2924, pp 2924–2929
prosthetic hand: analysis of user-prosthesis 43. Qiang S, Fang L, Findlay RD (2006)
interaction. IEEE Transactions on Robotics Generalized predictive control for a pneumatic
24(1):170–184. doi:10.1109/TRO.2007. system based on an optimized ARMAX model
910708 with an artificial neural network. International
32. Cipriani C, Controzzi M, Vecchi F et al (2008) conference on computational intelligence for
Embedded hardware architecture based on modelling, control and automation and inter-
microcontrollers for the action and perception national conference on intelligent agents, web
of a transradial prosthesis. 2nd IEEE RAS & technologies and internet commerce, Nov 28–
EMBS International conference on biomedical Dec 1 2006, pp 223–223
robotics and biomechatronics, 19–22 Oct 44. Haley P, Soloway D, Gold B (1999) Real-
2008, pp 848–853 time adaptive control using neural general-
33. Rodriguez-Cheu LE, Gonzalez D, Rodriguez ized predictive control. Proceedings of the
M () Result of a perceptual feedback of the American Control Conference, vol 4276,
grasping forces to prosthetic hand users. 2nd 1999, pp 4278–4282
IEEE RAS & EMBS international conference 45. Noriega JR, Wang H (1998) A direct adaptive
on biomedical robotics and biomechatronics, neural-network control for unknown nonlinear
19–22 Oct 2008, pp 901–906 systems and its application. IEEE Trans Neural
Netw 9(1):27–34
34. Pasluosta CF (2010) Nonlinear control strat-
egy for a cost effective myoelectric prosthetic 46. Norgaard M, Ravn O, Poulsen NK (2001)
hand. Dissertation, Lousiana Tech University, NNSYSID and NNCTRL tools for system
Ruston, Lousiana, USA identification and control with neural net-
works. Computing and Control Engineering
35. Pasluosta CF, Chiu AWL (2012) Evaluation of
Journal 12(1):29–36
a neural network-based control strategy for a
cost-effective externally-powered prosthesis. 47. Gunji D, Mizoguchi Y, Teshigawara S et al
Assist Technol 24(3):196–208 (2008) Grasping force control of multi-
fingered robot hand based on slip detection
36. Norgaard M, Ravn O, Poulsen NK, Hansen
using tactile sensor. IEEE international confer-
LK (2000) Neural networks for modelling and
ence on robotics and automation, 19–23 May
control of dynamic systems: a practitioner’s
2008, pp 2605–2610
handbook. Advanced textbooks in control and
signal processing. Springer, Great Britain 48.
Schuurmans J, Van Der Linde RQ,
Plettenburg DH et al (2007) Grasp force
37. Hornik K, Stinchcombe M, White H (1989) optimization in the design of an underactu-
Multilayer feedforward networks are universal ated robotic hand. IEEE 10th international
approximators. Neural Netw 2(5):359–366 conference on rehabilitation robotics, 13–15
38. Dahunsi OA, Pedro JO, Nyandoro OT (2009) Jun 2007, pp 776–782
Neural network-based model predictive con- 49. Kamikawa Y, Maeno T (2008) Underactuated
trol of a servo-hydraulic vehicle suspension sys- five-finger prosthetic hand inspired by grasping
tem. AFRICON ‘09, 23–25 Sep 2009, pp 1–6 force distribution of humans. IEEE/RSJ inter-
39. Haichen Y, Zhijun Z (2006) Predictive con- national conference on intelligent robots and
trol based on neural networks of the chemical systems, 22–26 Sep 2008, pp 717–722
Chapter 12
Application of Artificial Neural Networks

in Computer-Aided Diagnosis
Bei Liu
Abstract
Computer-aided diagnosis is a diagnostic procedure in which a radiologist uses the outputs of computer
analysis of medical images as a second opinion in the interpretation of medical images, either to help with
lesion detection or to help determine if the lesion is benign or malignant. Artificial neural networks (ANNs)
are usually employed to formulate the statistical models for computer analysis. Receiver operating charac-
teristic curves are used to evaluate the performance of the ANN alone, as well as the diagnostic perfor-
mance of radiologists who take into account the ANN output as a second opinion. In this chapter, we use
mammograms to illustrate how an ANN model is trained, tested, and evaluated, and how a radiologist
should use the ANN output as a second opinion in CAD.
Key words Artificial neural network (ANN), Medical image, Mammogram, CAD, Receiver operating
characteristic (ROC)
1 Introduction
The goal of computer-aided diagnosis (CAD) is to improve the

sensitivity, specificity, and efficiency of a radiologist’s diagnosis by
using computer analysis of medical images as a second opinion.
Using CAD to detect lesions and using CAD to classify lesions are
known as CADe and CADx, respectively.
Recently, Eadie et al. [1] reviewed 48 CAD studies (16 CADe
and 32 CADx) from 1992 to 2010. They concluded that, overall,
there is no clear benefit to using CADe, but CADx significantly
improves diagnosis in mammography and breast ultrasound, while
CADx shows an adverse impact on diagnosis based on lung CT
and dermatologic imaging. While the improvement of breast can-
cer diagnosis using CAD shows the great potential of the method,
the lack of evidence of benefit in some other applications indicates
that CAD is still in its infancy stage and more research and devel-
opment are needed in this field.
195
196 Bei Liu
Artificial neural networks (ANNs) have been widely used in

CAD for tasks such as pattern recognition and lesion classification
[2–6]. An ANN is a computational model loosely related to how
the human brain is presumed to operate; it can be viewed as a mul-
tivariate mathematical function that consists of interconnected
computational nodes (or neurons). The connection between a
neuron in one layer and the neurons in the preceding layer is
represented by a weighted linear combination of the output of
nodes in the proceeding layer, modified by a nonlinear activation
function (e.g., a sigmoid function). The weights are determined
through iterative ANN training, which minimizes the sum of the
squared error between the ANN output and the known target
values for training cases.
2 Materials
2.1 ANN Architecture Three-layer feedforward and error-backpropagation neural net-

works can approximate any multidimensional continuous function,
according to the universal approximation theorem [7], and this
ANN architecture is commonly used in CAD applications. The
number of input nodes equals the dimensionality of the input
feature space and, in the application described in this chapter, the
output layer has only a single node that outputs the likelihood of
malignancy of a known lesion. In a CAD application, the number
of hidden nodes is limited for better generalizability because the
number of training cases is finite [8, 9]; this number is determined
empirically based on the dimensionality of the input feature space
and the size of the training dataset. Once the number of hidden
nodes has been chosen, the mathematical function format that the
ANN represents is also delimited, and it is no longer true that the
ANN can approximate any arbitrary multidimensional continuous
function.
For the ANN used by Rana et al. in the CAD study based on
mammograms, eight input features were available, and thus eight
input nodes were used, together with six hidden-layer nodes [10].
The architecture of such an ANN is illustrated in Fig. 1.
2.2 Data Collection An ANN is actually a mathematical model with many weight
parameters. For the ANN architecture in Fig. 1, there are 54
weight parameters that need to be optimized in the ANN training;
the number of training cases should be much larger than the num-
ber of parameters to ensure the generalizability of the ANN model.
Therefore, medical images for a large number of patients need
to be collected, in order to build an ANN model for CAD applica-
tion. The patients are partitioned into three groups: training cases,
validation cases, and testing cases, to ensure generalizability of the
ANN. For a patient with multiple images, such as a breast MRI and
ANN Application in CAD 197
Fig. 1 8-6-1 ANN architecture. There are 54 internode connections, and thus 54
weight parameters
a mammogram, or a CC-view mammogram and an MLO-view

mammogram, the data are considered as one case; therefore, all the
images of this patient should be in just one of the training group,
the validation group, and the testing group. The data cannot be
partitioned such that, for example, the CC-view mammogram of a
patient is in the training group while the MLO-view mammogram
of the same patient is in the validation or the testing group.
The training cases are used to optimize the weight parameters
through minimization of the cost function, which is usually the
sum of the squared error between the ANN outputs and the train-
ing target. The validation cases are used to find the best perform-
ing ANNs among the set of ANNs trained with different learning
rates/momentum, and among ANNs trained to a different num-
ber of epochs, etc. The testing cases are used to evaluate the ANN
performance independently.
2.3 Performance The receiver operating characteristic (ROC) is used to evaluate the
Evaluation overall performance of the trained ANN model and the CAD sys-
tem. An ROC curve completely describes the trade-off between
sensitivity and specificity at different operating points [11, 12],
where sensitivity or true positive fraction (TPF) is defined as the
probability that an actually positive case is diagnosed as positive
and specificity is defined as the probability that an actually negative
case is diagnosed as negative. More often, the false positive fraction
(FPF) is used, which is defined as the probability that an actually
negative case is diagnosed as positive. One can easily see that, by
definition, FPF = 1–specificity. The plot of TPF vs. FPF is the ROC
curve (Fig. 2). The area under an ROC curve (AUC) is a very use-
ful summary index to describe the CAD system’s overall perfor-
mance: AUC can be viewed as the average sensitivity or average
specificity of the CAD system.
198 Bei Liu
Fig. 2 Three ROC curves. AUC is the average sensitivity or average specificity
Fig. 3 A binormal distribution of negative cases and positive cases
A binormal model has been successfully used to fit data in a

wide variety of practical situations [13, 14]; this assumes that the
latent decision variable can be transformed to a pair of normal dis-
tributions monotonically: a standard normal distribution for actu-
ally negative cases, and a normal distribution with mean of a/b and
standard deviation of 1/b for actually positive cases (Fig. 3). Under
a binormal model, the AUC value is usually called Az value.
Free ANN software can be downloaded from many websites,
which can be found using a web search; professional ANN pack-
ages are available from Matlab or SAS; these require payment of a
license fee. ROC analysis software can be downloaded from URL
http://metz-roc.uchicago.edu/MetzROC/software free by setting
up a user account.
3 Methods
3.1 Medical Image 1. After collecting enough medical images (digital or film based)
Processing for patient cases to build an ANN model for CAD applications,
all the film-based images need to be digitized and stored in a
CAD server along with the digital images.
2. Image pre-processing is often necessary to remove noise and
artifacts in the images. Histogram equalization is applied on
each image to improve contrast.
3. Regions of interest (ROIs) are defined so that the ROIs, which
are simply rectangular regions, enclose the lesions. It is possi-
ble that there will be multiple ROIs on one medical image.
4. Lesions are segmented automatically using computer software.
Depending on the applications, many segmentation methods
can be used: from simple thresholding to region growing,
watershed, or an active contour model (snake). For microcalci-
fication lesions of the breast, the shape and size features
strongly depend on the segmentation algorithm used [15, 16];
therefore, various segmentation methods should be tried and
tested in order to choose a segmentation method that most
effectively fits the CAD application.
3.2 Feature 5. Feature extraction: Based on the lesion segmentation, lesion

Extraction and Feature features such as shape and size can be calculated. Feature
Selection extraction is a form of dimension reduction of the input data,
which are the raw medical images in a CAD application.
Feature extraction is essential in order to reduce computation
time, and, more importantly, to avoid overfitting in ANN
training. One convenient way to accomplish feature extraction
is to quantify the radiologists’ perception in image reading.
For example, based on radiologists’ experience, Jiang et al. [4]
used eight features to design a CAD system for the classifica-
tion of microcalcifications of the breast: cluster circularity, clus-
ter area, number of microcalcifications, average effective
volume of microcalcifications, average area of microcalcifica-
tions, second highest microcalcification-shape-irregularity
measure in a cluster, relative standard deviations in effective
thickness, and volume.
6. Nonimage features such as patient age can also be used to
build the ANN model [17, 18].
7. Feature selection: Even after feature extraction, which is a form
of dimension reduction of the input image data, the dimen-
sionality of the feature space needs to be reduced to avoid
overfitting. More importantly, multicollinearity, in which two
or more features are highly correlated, must be removed, in
order to train the ANN more effectively and to obtain a more
200 Bei Liu
stable ANN model [19]. Another goal of feature selection is to

remove irrelevant features, which have no predicting power.
Stepwise feature selection can use either bottom-up or
top-down methods. Bottom-up methods start with an empty
feature space, and then add one feature at a time to build an
ANN or a linear discriminant analysis (LDA) model. If, when a
feature is added, model performance is better than a preset
criterion, then this feature is retained; otherwise, it is discarded;
this procedure is repeated until all the features have been tried.
On the other hand, the top-down method starts with all the
features; one feature is removed and an ANN or an LDA model
is constructed. If the model performance is not too much
worse based on a preset criterion, then this feature is discarded;
otherwise it is retained. For large training datasets, LDA should
be used for feature selection [20] since it is fast when handling
large datasets; when the training dataset is small, ANN is used
for feature selection.
3.3 ANN Training 8. The image dataset is partitioned into three sub-datasets for
ANN modeling: a training dataset, a validation dataset, and a
testing dataset. The goal of ANN training, which is an iterative
process, is to optimize the weight of each nodal connection to
minimize the cost function. The cost function is the sum of the
squared error between ANN outputs and training target val-
ues, which is defined from the golden truth based on biopsy. In
the task of classifying a breast lesion as benign or malignant,
the target values for benign cases and malignant cases are set to
be 0 and 1, respectively (see Notes 1 and 2).
9. Backpropagation is used for ANN training. In this method the
weight correction is ∆wk(i) = −η∂E/∂wk + α∆wk(i − 1) in the i-th
iteration, where w is the weight factor, E is the cost function,
and ∂E/∂wk is the gradient descent force that drives the cost
function to the minimum. The parameter η is the learning rate;
this controls how rapidly the ANN training converges, and the
parameter α is the momentum, which is used to reduce the
weight fluctuation. The optimal values of η and α are problem
dependent and are determined by trial and error for a specific
problem [21].
10. Since backpropagation is a gradient descent method, there is a
danger that the ANN training might get trapped in a local
minimum. It is generally impossible to be certain that the
global minimum has been located, so various initial weights
should be tried for ANN training in order to obtain a good
ANN model. The area under the ROC curve (AUC) is calcu-
lated from the validation dataset for each epoch, or every ten
epochs, and the ANNs with the maximum AUC values are
tested using the testing dataset. Since the testing dataset is
not involved in the ANN training or ANN selection, it is
completely independent, and the ROC curve calculated from

the testing dataset is an independent evaluation of the ANN
performance (see Notes 3 and 4).
3.4 ROC Analysis 11. ROC analysis is used to evaluate the ANN performance and
and Observer Study LABROC4 is used for binormal ROC curve fitting [12, 22].
LABROC4 and some other ROC analysis program such as
CORROC, which calculates the p value for two ROC curves,
are developed by the University of Chicago and can be down-
loaded free from http://metz-roc.uchicago.edu/MetzROC/
software. The 95 % confidence interval for Az, which is actually
the AUC after binormal fitting, is also calculated.
12. An observer study is conducted to evaluate the performance of
the CAD. A number of radiologists should be recruited for an
observer study since the CAD performance is radiologist
dependent. Before the observer study, each radiologist is
trained so that they are familiar with the CAD system, includ-
ing how to use the software, and to become familiar and com-
fortable with the setting and the flow of the study. The
radiologists are given some details of how the ANNs are
trained, such as the image features and nonimage features that
are used for ANN training, the size of the training dataset, the
validation dataset and the testing dataset, as well as the Az
value of the ANN model applied on the testing case and its
standard error. The observers also read a number of cases with
and without computer aid, and compare those data with the
known outcomes, so that they can build up adequate confi-
dence in use of the CAD system. After this brief training,
observers are asked to view a set of medical images, which can
be the ANN testing images or other independent images, but
cannot be the training or validation images, with or without
computer-estimated likelihood of malignancy.
For example, in the CAD observer study conducted by Rana
et al. [10], four radiologists were asked to view a set of mam-
mograms with microcalcification lesions with and without com-
puter aid. Each radiologist drew a square ROI to enclose the
microcalcification lesion and the computer algorithm automati-
cally calculated the number of calcifications and segmented
them. For 93 out of 332 cases, observers were also asked to
place the number of calcifications into one of the four catego-
ries to help the computer algorithm set a threshold for calcifica-
tion identification [23]: (a) 3–5, (b) 6–10, (c) 11–30, or (d) 31
or more calcifications. The Az value of the ANN model used was
0.79 with a standard error of 0.05. With and without computer-
estimated likelihood of malignancy, each observer read both the
CC-view and MLO-view mammogram for each patient and
assigned a BI-RADS category for each patient: benign findings,
probably benign, suspicious abnormality, and highly suggestive
202 Bei Liu
of malignancy. The other categories are not used because all the
mammograms have calcification lesions. ROC analysis was then
performed and compared with the radiologists’ diagnosis with
and without computer aid (see Notes 5 and 6).
4 Notes
1 When to stop ANN training: Early stopping is often used to

avoid overfitting as a form of regularization. However, the
problem with this approach is that the question of just when
training should stop is not a trivial one; consequently, early
stopping has often been employed in an ad hoc fashion. Our
experience is to calculate and monitor the ANN performance
vs. epoch using an independent validation dataset and use this
to help choose the best epoch to stop ANN training, even
though this method leaves a smaller dataset for ANN training.
2 It is crucial to collect a big dataset in order to develop a CAD
system for clinical use. However, in the research phase, it may
be too expensive, or even impossible, for a research group to
collect a dataset of suitable size; thus small datasets were often
used to test new CAD ideas, such as a new segmentation
method, a new image feature, or a new imaging modality.
In this situation, a round-robin method (also known as a leave-
one-out method) is often used [4]. In this method, all but one
of the cases are used for ANN training and the one case left out
is used for testing; thus this test case is independent of the
training cases. This training and testing procedure is repeated
until each case has been left out once; therefore each case has
an independent test score and ROC analysis can be applied to
these scores to obtain the overall ANN performance.
3 Ideally, the ANN output is an estimation of the posterior
probability, provided that the training dataset is sufficiently
large and the ANN architecture is sufficiently complex [24, 25].
Therefore the ideal ANN output has zero variance. However,
in reality the ANN output has finite variance because the
training data size is finite. In our application, the goal of ANN
training is to achieve a high AUC under the ROC curve and
low variability in the ANN outputs. However, if we trained
several ANNs with similar overall performance as measured by
the AUC, a significant variability in ANN outputs is actually
beneficial: it indicates that there is significant room to improve
the overall ANN performance, and it can be simply achieved
by taking the average, the maximum, or the minimum. So one
trick to achieve a higher performance classifier is to train sev-
eral ANN with different methods and then combine the
results. For example, one can train an ANN for each of the
segmentation methods (region-growing, watershed, or active

contour model) and then combine the ANN outputs in order
to achieve higher overall ANN performance.
4 When the ANN training truth provides richer information than
merely benign/positive, this extra information can sometimes
be used in ANN training if the dataset is small. For example, the
golden truth of ANN training for breast cancer CAD is based
on the biopsy result: training targets of 0 and 1 are assigned to
benign and malignant cases, respectively. However, the biopsy
provides more information than benign/malignant. For benign
cases, two histology subtypes exist: benign and benign with
higher risk; for malignant cases, there are four histology sub-
types: noncomedo ductal carcinoma in situ (DCIS), comedo
DCIS, DCIS with microinvasion, and invasive carcinoma. This
sequence of histology subtypes is one possible representation of
an ordered spectrum from benign tissue at one extreme to
aggressive cancer at the other [26]. Ordered ANN training tar-
get values can be assigned according to the histology subtypes
in ANN training for small training datasets [25].
5 In the mammogram study [10], eight input features were used
and an 8-6-1 ANN architecture was employed. However, if the
amount of training data is large, more hidden nodes can be
used to increase the ANN complexity, and thus enhance its
predictive ability.
6 In the breast cancer CAD based on mammograms, each patient
will have two ANN scores that correspond to the probabilities
of malignancy calculated from CC- and MLO-view mammo-
grams [27–29]. A natural way to use this data is to calculate the
average score for the patients. As expected, averaging improves
the AUC under the ROC curve compared to single-view images
[30]; however, averaging is not always the optimal way of com-
bining the two scores: depending on the standard deviation σ of
the distribution of positive cases after fitting to a binormal
model and the correlation of the two ANN scores, taking the
maximum score or taking the minimum score can outperform
the averaging approach in some situations [31, 32].
References
1. Eadie LH, Taylor P, Gibson AP (2012) A sys- benign microcalcifications on mammograms:
tematic review of computer-aided diagnosis in texture analysis using an artificial neural net-
diagnostic cancer imaging. Eur J Radiol 81: work. Phys Med Biol 42:549–567
e70–e76 4. Jiang Y, Nishikawa RM, Schmidt RA et al
2. Baker JA, Kornguth PJ, Lo JY et al (1996) (1999) Improving breast cancer diagnosis with
Artificial neural network: improving the quality computer-aided diagnosis. Acad Radiol 6:
of breast biopsy recommendations. Radiology 22–33
198:131–135 5. Huo Z, Giger ML, Vyborny CJ et al (2002)
3. Chan H, Sahiner B, Petrick N et al (1997) Breast cancer: effectiveness of computer-aided
Computerized classification of malignant and diagnosis—observer study with independent
204 Bei Liu
database of mammograms. Radiology 224: 19. Matignon R (2005) Neural network modeling
560–568 using SAS enterprise miner. AuthorHouse,
6. Suzuki K, Li F, Sone S et al (2005) Computer- Bloomington
aided diagnosis scheme for distinction between 20. Way TW, Sahiner B, Hadjiiski LM et al (2010)
benign and malignant nodules in thoracic low- Effect of finite sample size on feature selection
dose CT by use of massive training artificial and classification: a simulation study. Med Phys
neural network. IEEE Trans Med Imaging 37(2):907–920
24:1138–1150 21. Rojas R, Feldman J (1996) Neural networks: a
7. Richard HH, Lippman RP (1991) Neural net- systematic introduction. Springer, New York
work classifiers estimate Bayesian a posteriori 22. Metz CE, Herman BA, Shen JH (1998)
probabilities. Neural Comput 3:461–483 Maximum likelihood estimation of receiver
8. Swingler K (1996) Applying neural networks: a operating characteristic (ROC) curves from
practical guide. Academic, London continuously distributed data. Stat Med 17:
9. Berry MJA, Linoff G (1997) Data mining tech- 1033–1053
niques. Wiley, New York 23. Salfity MF, Nishikawa RM, Jiang Y et al
10. Rana RS, Jiang Y, Schmidt RA et al (2007) (2003) The use of priori information in the
Independent evaluation of computer classifica- detection of mammographic microcalcifica-
tion of malignant and benign calcifications in tions to improve their classification. Med Phys
full-field digital mammograms. Acad Radiol 30:823–831
14:363–370 24. Rojas R (1996) A short proof of the posterior
11. Swets JA (1978) ROC analysis applied to the probability classifier neural networks. Neural
evaluation of medical imaging techniques. Comput 8:41–43
Invest Radiol 14:109–121 25. Liu B, Jiang Y (2013) A multitarget training
12. Metz CE (1986) ROC methodology in radio- method for artificial neural network with appli-
logic imaging. Invest Radiol 21:720–733 cation to computer-aided diagnosis. Med Phys
13. Swets JA (1986) Forms of empirical ROCs in 40:011908. doi:10.1118/1.4772021
discrimination and diagnostic tasks: implication 26. Rosen PP (2001) Rosen’s breast pathology.
for theory and measurement of performance. Lippincott Williams & Wilkins, Philadelphia
Psychol Bull 99:181–198 27. Jiang Y, Nishikawa RM, Wolverton DE et al
14. Hanley JA (1988) The robustness of the binor- (1996) Malignant and benign clustered micro-
mal assumption used in fitting ROC curves. calcifications: automated feature analysis and
Med Decis Making 8:197–203 classification. Radiology 198:671–678
15. Veldkamp WJH, Karssemeijer N (1996) 28. Chan H, Sahiner B, Lam KL et al (1998)
Influence of segmentation on classification of Computerized analysis of mammographic
microcalcifications in digital mammography. microcalcifications in morphological and tex-
IEEE engineering in medicine and biology ture feature spaces. Med Phys 25:2007–2019
society, bridging disciplines for biomedicine, 29. Huo Z, Giger ML, Vyborny CJ (2001)
proceedings of the 18th annual international Computerized analysis of multiple-
conference of the IEEE 3, pp 1171–1172 mammographic views: potential usefulness of
16. Paquerault S, Yarusso LM, Papaioannou J et al special view mammograms in computer-aided
(2004) Radial gradient-based segmentation of diagnosis. IEEE Trans Med Imaging 20:
mammographic microcalcifications: observer 1285–1292
evaluation and effect on CAD performance. 30. Metz CE, Shen J (1992) Gains in accuracy
Med Phys 31:2648–2657 from replicated reading of diagnostic images:
17. Shi J, Sahiner B, Chan HP et al (2008) prediction and assessment in terms of ROC
Characterization of mammographic masses analysis. Med Decis Making 12:60–75
based on level set segmentation with new 31. Liu B, Metz CE, Jiang Y (2004) An ROC com-
image features and patient information. Med parison of four methods of combining
Phys 35(1):280–290 information from multiple images of the same
18. Singh S, Maxwell J, Baker JA et al (2011) patient. Med Phys 31:2552–2563
Computer-aided classification of breast masses: 32. Liu B, Metz CE, Jiang Y (2005) Effect of cor-
performance and interobserver variability of relation on combining diagnostic information
expert radiologists versus residents. Radiology from two images of the same patient. Med Phys
258:73–80 32:3329–3338
Chapter 13
Developing a Multimodal Biometric Authentication

System Using Soft Computing Methods
Mario Malcangi
Abstract
Robust personal authentication is becoming ever more important in computer-based applications. Among
a variety of methods, biometric offers several advantages, mainly in embedded system applications. Hard
and soft multi-biometric, combined with hard and soft computing methods, can be applied to improve the
personal authentication process and to generalize the applicability.
This chapter describes the embedded implementation of a multi-biometric (voiceprint and finger-
print) multimodal identification system based on hard computing methods (DSP) for feature extraction
and matching, an artificial neural network (ANN) for soft feature pattern matching, and a fuzzy logic
engine (FLE) for data fusion and decision.
Key words Artificial neural network, Fuzzy decision logic, Soft-biometric data, Multi-biometric
1 Introduction
More and more of the devices we deal with every day depend on
embedded systems. This trend can reasonably be expected to accel-
erate going forward. Because of this, as well as due to other factors,
the security of systems and of data is likely to continue to present
challenges that developers will attempt to meet by devising new
tools. There can be little doubt that many such methods will rely
on the biometric approach [1–4].
Biometrics is uniquely suited to authentication tasks in embed-
ded systems because it offers simultaneous solutions to the twin
problems of the meager demand for resources that we wish to
place on such systems and of the high security requirements that
their potential applications can be expected to mandate [5, 6]. It is
not hard to imagine wanting to limit demand on resources if we
start from the very basic concept of a system that has no keyboard,
is designed to have the smallest possible footprint, and can be
considered subsidiary to a larger system, machine, or process. If we
consider the embedded system’s role as gatekeeper to a much
205
206 Mario Malcangi
larger and more complex reality, say the electronic ignition key to
a sophisticated, automated installation, it is not hard, either, to
envision the degree of security requirements for which a minuscule
device is ultimately responsible. As such devices proliferate, the
demand for lightweight yet robust authentication methods can
only grow.
Because biometrics depends on physiological traits or on dis-
tinctive behavior [7, 8] unique to the individual whose identity is
to be authenticated, or—indeed—a combination of the two, it can
rightly stake a claim to being superior, at least potentially, to cur-
rent and established authentication methods like personal identifi-
cation numbers (PINs), passwords, or smart cards. The individual’s
biometric data is always available, is nonrandomly unique, cannot
be transferred to another party, cannot be forgotten, is not subject
to theft, and cannot be guessed. These advantages mean that bio-
metrics provides very high security compared to traditional identi-
fication methods that rely on the possession of a token, such as a
card, key, or chip, or of personal knowledge, as in the case of a
password [9].
Despite the superiority of biometric authentication due to such
advantages and despite its rapid spread in microelectronics, mass-
market adoption has lagged. The primary reason widespread appli-
cation of biometrics has not taken off is that it does not offer
surefire authentication the way a password does. A password can
always perform. On the other hand, biometric features, in their
original form, are analog information. As a result, they are subject
to variation when captured by a biometric scanning device. This
applies to fingerprint readers, microphones for voice-pattern rec-
ognition, cameras for facial feature matching, and any other digital
system that has to match measured, analog, human traits. Of course
such features can be digitized, but the data will nonetheless have
been processed through fuzzy logic. Fuzziness can only be mini-
mized so much, given that the original, analog features are, them-
selves, inherently fuzzy.
False acceptance rate (FAR) and false rejection rate (FRR) [10]
can, of course, be minimized with classical pattern-matching tech-
niques [11–13] but a 100 % correct acceptance rate cannot be
achieved this way. However, the intrinsic fuzziness of biometric
features makes it logical to look to soft-computing approaches
[14–16] in the design of processes that match biometric feature
patterns in the hope that nearly 100 % correct authentication might
be attained that way. For example, even in prohibitive conditions,
human beings always manage to recognize a familiar face. The
human mind processes biometric features fuzzily and neurally.
Soft-computing data are variations and uncertainties. Biometric
features vary constantly, present challenges to analytical description,
and inherently fall short of belonging to their owner 100 % of the time.
Biometric Authentication 207
It follows that soft-computing methods ought to work well for

measuring and matching biometrics. Moreover, what humans do
to reach nearly 100 % authentication rates amounts to acquiring
data from multiple sources. They combine speech traits and facial
features when matching a biometric identity to an individual [17].
In other words, they carry out neural and fuzzy processing on
aggregated biometric feature sets.
While biometric authentication has seen surprisingly slow
adoption generally, this has proved even more the case with the
application of traditional biometric solutions to embedded systems
[18]. Because embedded systems have limited resources—smaller
memory and slower processors—they are ill suited to hungry
authentication applications. Consequently, current installations of
embedded systems typically rely on PIN codes and the like.
However, an authentication method based on soft biometric
[19] data has the potential to be more miserly with computing
resources than hard biometrics, using less data to map the identity
of the person to be authenticated. Furthermore, multiple-criteria
biometric authentication processes [8, 20, 21] can be devised so as
to match the input capacities of embedded systems. By combining
soft-computing methods and multi-biometrics, we can optimize
such authentication for implementation on embedded systems.
One approach that holds promise in this regard involves set-
ting up authentication based on a combination of voiceprint and
fingerprint that uses hard-computing digital signal processing to
extract features and match them but then turns to an artificial neu-
ral network (ANN) [22] for soft feature pattern matching and to a
fuzzy logic inferential engine (FLE) for data fusion and decision
making. Such a setup can be designed to provide highly robust,
personal, biometric authentication. Experiments have been carried
out where this can be accomplished with a single-chip, floating-
point, digital signal processor.
2 Materials
To design a multimodal soft computing-based hard/soft biometric

embedded system several methodologies and technologies occur:
●● Biometric sensors.
●● Sensor data acquisition.
●● Feature extraction algorithms.
●● Hard computing pattern matching.
●● Soft computing pattern matching (artificial neural network).
●● Soft computing fusion and decision (fuzzy logic).
●● System modeling and simulation environments.
208 Mario Malcangi
2.1 Biometric The biometric sensors are specifically designed to capture physiologic
Sensors or behavioral signals generated by human beings. The sensor is the
first device of the signal chain. It captures and converts a physical
property into analog signals (commonly electrical) or digital sig-
nals (when it embeds the mixed-signal electronics).
2.1.1 Voiceprint Sensors The voiceprint sensor is a voice-grade microphone system able to
capture the utterance, preserving the intelligibility of the speech.
●● Single microphone: captures the acoustic wave of the utterance
as a direct sound source (position sensitive; sensitive to sur-
rounding noise).
●● Dual microphone: captures the acoustic sound wave of the
utterance and the surrounding audio noise (partially position
sensitive; low sensitivity to surrounding noise).
●● Microphone array: captures the acoustic sound wave of the
utterance by a set of microphones spatially distributed (posi-
tion insensitive; highly insensitive to surrounding noise).
A single microphone has been used in this design (see Note 1).
2.1.2 Fingerprint Sensors A fingerprint sensor is an electronic device capable to capture the
image of the fingerprint pattern and make it available as a bit map.
Several technologies have been applied to develop fingerprint
sensors:
●● Optical (special digital camera): captures visible light emitted
from a phosphor layer which illuminates the surface of the fin-
ger—advantages: noncontact, not electrostatic discharge sensi-
tive—disadvantages: capabilities affected by the quality of skin,
easily fooled.
●● Ultrasonic (based on medical ultrasonic imaging principles):
captures the images of a fingerprint by penetrating the epider-
mal layer of skin with high-frequency sound—advantages:
noncontact, not electrostatic discharge sensitive, capabilities
not affected by the quality of skin, very difficult to fool—disad-
vantages: the price is significantly higher than for a capacitive
fingerprint scanner.
●● Capacitance (based on electrical capacitance): consists of an
array of capacitors in which one of the two plates is the dermal
layer and the dielectric is the epidermal
–– Passive: Capacitance is measured across each sensor capaci-
tor to form a pixel of the fingerprint image.
–– Active: A voltage is applied to the skin first, and then the
charge of each capacitor of the array is measured to form a
pixel of the fingerprint image.
Advantages: compatible with microelectronic integration—

disadvantages: electrostatic discharge sensitive.
A passive capacitive sensor has been used in this design
(see Note 2).
2.2 Sensor Data Sensor data acquisition concerns the conditioning and digital rep-
Acquisition Chain resentation of the voiceprint and fingerprint information captured
by the respective biometric sensors. Conditioning is necessary if
the sensor is affected by nonlinearities and if it is not dynamically
compatible with analog-to-digital conversion (ADC) subsystem.
2.2.1 Microphone Sensor The microphone sensor data acquisition process is sensitive to several
Data Acquisition nonlinearities and mixed signal constraints. The microphone chain
path needs conditioning and digital-to-analog adaptation, so opti-
mal data can be available at the processing stage:
●● Transducer linearization.
●● Automatic gain control.
●● Filtering.
●● Sampling.
●● Quantization.
The microphone nonlinearities need to be compensated to
avoid distorted measurements at the level of feature extraction.
This can be done in the analog domain by electronic circuitry
tuned on the specific microphone, or in the digital domain at the
preprocessing stage of the signal acquisition. In this design the
second option has been applied because it is more flexible and
adaptive. The microphone transfer function is estimated at calibra-
tion time, and then the nonlinearity compensated for by applying
at run time the inverse function to the captured signal.
The amplification is required because the small signals from
the microphone need to be adapted to the analog-to-digital con-
verter (ADC) input dynamic range to have higher signal-to-
quantization-noise ratio (SQNR).
Low-pass filtering (antialiasing) and high-pass filtering (offset
removal) are applied prior to sampling the microphone signal to
ensure lowest frequency distortion of the pulse code modulated
(PCM) stream to be quantized and optimal SQNR.
2.2.2 Fingerprint Sensor The solid-state capacitive fingerprint sensor integrates all the required
Data Acquisition conditioning and digitalization resources so that optimal digital fin-
gerprint image data is available at its output. Fingerprint image is
captured at 513 dpi (dot per inch) resolution as a bitmap image
consisting of 64,512 pixels (224 horizontal and 288 vertical).
Each pixel has 8-bit-depth quantization level (gray-level images).
The fingerprint digital image array acquisition process is completed
by synchronous serial transfer to the application processor (DSP).
210 Mario Malcangi
2.3 Feature To extract the features, a set of digital signal processing algorithms
Extraction Algorithms is applied to the captured utterance.
2.3.1 Voiceprint Hard The following formulae have been applied to measure the voice-
Features print hard features [23]:
●● Root mean square (RMS):
N -1
1
RMS j =
N
ås (m ) j
2
(1)
m =0

●● Zero-crossing rate (ZCR):
N -1
ZCR j = å 0.5 sign ( s (m ) ) - sign(s (m - 1)
j j (2)
m =0
●● Autocorrelation (AC):
N N +1-i
AC j = å å s ( j ) s (i + j - 1 )
j j (3)
i =1 j =1

●● Cepstral linear prediction coefficients (CLPC):
m -1
ækö
CLPC j = am + å ç ÷ c k am -k (4)
k =1 è m ø
with
c0 = r(0): first autocorrelation coefficient
am: prediction coefficients
sj(n) = wN(n)s : j-th windowed part of the uttered stream s
æ 2p n ö
wN (n ) = 0.54 - 0.46 cos ç ÷ : N-length weighting Hamm-
èN -1ø
ing window
The above are short-time measurements executed by multiply-
ing the N samples-wide weighting Hamming window w(n) by the
uttered speech stream s(n).
RMS is the root mean square of the windowed speech segment.
Such measurement helps to identify phonetic unit end-points.
ZCR is the time-domain measurement of the dominant fre-
quency information. It is used to determine whether or not the
current processed uttered speech segment is voiced or unvoiced
and also to find the frequency (band) with the major energy
concentration.
AC is the measurement of the speech-pitch frequency. It
preserves information about pitch-frequency amplitude while

ignoring phase. Phase is unimportant for the purpose of speech
identification.
CLPC is the LPC-Cepstral feature vector that models the
vocal tract.
2.3.2 Voiceprint Soft The following soft features are extracted from speech:
Features ●● Speed.
●● Stress.
Speed is measured as the total duration of the speech utterance.
Stress is measured as the ratio between the peak amplitude of the
stressed vowel and the average amplitude of the whole utterance.
Both these voiceprint features are related to the way the person
is used to speaking a requested word.
2.3.3 Fingerprint Hard Several preprocessing steps are executed on the captured grayscale
Features fingerprint image from the stage of bitmap to its minutiae-based
representation:
●● Orientation field and region of interest.
–– Normalization of the captured fingerprint image.
–– Image segmentation in blocks.
–– x and y gradient computation for each pixel in each block.
–– Local orientation for each pixel.
–– Orientation field correction.
●● Positioning (delta and core localization).
●● Ridging.
●● Ridge thinning.
Fingerprint hard features are the minutiae. The criteria used
for minutiae extraction from the thinned ridge image are the
following:
●● If a ridge pixel has two or more 8-nearest, then it is a
bifurcation.
●● If a ridge pixel has only one 8-nearest, then it is a
termination.
As the crest-valley image is two levels encoded (1-0), the map-
ping algorithm is
7
if åp
i =0
i >2
the pixel is B (5)

else
the pixel is T
where pi are the 8-nearest pixels of the pixel to be classified as

bifurcation B or termination T.
The most critical step in this procedure is to avoid computing
false minutiae caused by noise in the scanned fingerprint image.
212 Mario Malcangi
To overcome this problem, a backtracking control is executed on

each feature pattern before it is validated, to check that each of the
three branches of the bifurcation is significantly long:
●● Starting from the bifurcation all the three paths are checked,
stopping if more than k pixels or a termination pixel is detected.
●● If the stop is due to the termination detection, then the bifur-
cation and the terminations are invalidated.
The scanned fingerprint image is transformed into a set minu-
tiae encoded by its x, y coordinates and the direction of the ridge
corresponding at that position.
2.3.4 Fingerprint Soft Two fingerprint soft features are extracted from fingerprint cap-
Features tured image:
●● Total area.
●● Mean intensity.
Both these two fingerprint features are related to the way the
person approaches contact with the fingerprint sensor.
The total area is measured as the ratio between the total pixels
available on the fingerprint scanning device and the total pixels of
the captured fingerprint image that have a value higher than an esti-
mated peak noise level. The peak noise level is estimated at calibra-
tion time and it is applied at run time (enrolling and identification).
The mean intensity is measured as average intensity of all the
pixels with a value higher than a threshold level. The threshold
level is 10 % higher than the peak noise level.
2.4 Hard Computing The captured fingerprint and voiceprint hard features have to be
Pattern Matching matched with enrolled features (templates).
2.4.1 Voiceprint Hard Two methods are applied to score the person’s identity. One is
Features based on measuring Mahalanobis distance, and the other on mea-
suring the distance of dynamic time warping—k-nearest neighbor
(DTW-KNN).
The Mahalanobis distance measurement is
Di ( x ) = ( x - x ) W -1 ( x - x )
T
(6)

where W is the covariance array computed using the average and
the standard deviation features of the utterance. The input pattern
x is processed with reference to the utterance-averaged feature vec-
tor x that represents the person to be identified. The distance
Di(x) is a score for the authorized user.
The DTW-KNN (dynamic time warping—k-nearest neighbor)
method combines the dynamic-time-warping measurement with the
k-nearest neighbor decision algorithm. The DTW clusters similar
template
search space
pattern
m
M
Warp
1
input
pattern
1 N
Fig. 1 DTW boundaries for voiceprint matching
elements that refer to a feature into classes (Fig. 1). The cost

function is computed using Euclidean distance, with a granularity of
one frame. The KNN algorithm is then applied to select k minimal
distance matching and to choose the most recurring person in k
minimal distance matches. This results in lower false-positive and
false-negative rates during identification, compared to the original
DTW algorithm.
2.4.2 Fingerprint Hard Fingerprint hard feature pattern matching is based on minutiae.
Features Pattern matching consists of a procedure that first tries to align the
template pattern and the input pattern, and then computes an
overlapping score (Fig. 2). The score is a measurement of the
authenticity of the person who enrolled the input.
2.5 Soft Computing The captured fingerprint and voiceprint soft features have to be
Pattern Matching matched with enrolled features (templates). This is done using the
artificial neural network (ANN) soft computing paradigm.
2.5.1 Artificial Neural The ANN is a signal processing and pattern classification paradigm
Network (ANN) inspired by the structure and functions of biological neural net-
works. Information signals that flow through the ANN modify the
connections, enabling the learning process.
214 Mario Malcangi
∆ search space
template
input
∆q
∆
reference
Fig. 2 Boundaries for fingerprint matching
The ANN applied to classify the soft biometric features is a

feed-forward back-propagation neural network (FFBP-ANN) par-
adigm because of the behavioral nature of this kind of biometric
data (Fig. 3a). It is a three-layered parallel processing network that
processes the input data patterns at the lower layer, matches the
data at the middle layer, and organizes the results at the upper
layer. Its input nodes are fully connected to all the nodes in the
hidden layer, and the hidden layer is fully connected to the output
nodes. In this network the feed-forward action consists in that the
i node at l + 1th layer receives signals from the j node in the lth
layer conditioned via weight wijl. The activation of the node i at the
l + 1th layer is given by
æM ö
f çç åwijl x lj ( k ) - Ji ÷÷ (7)
è j =1 ø
This is the activity of the ith node at the layer l + 1 for the kth input
xj processed by the ANN, assuming that M nodes are in the lth
layer.
Input and output layers have a linear activation function that
controls the connection. A nonlinear (sigmoid) activation function
(a)
x...0 w0 f(x;w; θ)
f(x;w; θ) x...j wj
wM
xj wj xM θ
θ
(b)
(b)
x...0 w0 f(x;w; θ)
xj wj
... wM
xM θ
(c)
Fig. 3 (a) FFBP-ANN for voiceprint and fingerprint soft feature classification, (b) linear activation function for
input and output layers, and (c) sigmoid activation function for hidden layer
(Fig. 3b) connects hidden-layer nodes to output-layer nodes

according to the following formulae:
1
si =
1 + e- b Ii (8)
I i = åwij x j - Ji
j

where ϑ is the activation threshold of the ith node and β is a con-
stant that controls the slope of the semi-linear region of the sig-
moid. When β is very small, the sigmoid approximates the linear
activation function (Fig. 3c), so the same function can be applied
to input, hidden, and output layers by choosing appropriate values
for β constant.
2.6 Soft Computing The fusion and decision logic is based on the fuzzy logic inferential
Fusion and Decision paradigm. A fuzzy logic engine processes the crisp inputs (scores
and features), applies to them a set of inferential rules, and makes
the fuzzy decision. This is done by the fuzzy logic engine.
2.6.1 Fuzzy Logic Engine The fuzzy logic engine (FLE) used to implement the data and deci-
sion fusion and to infer about the final decision is a zero-order
Sugeno-type (Fig 4a). It fuzzyfies the inputs (soft features and scores)
216 Mario Malcangi
a
µ µ µ K1 K2 K3
A1 A2 A3 B1 B2 B3
1 1 1
0.8
OR 0.8
0.1 (max)
0 0 0
x1 X y1 Y k2 K
Rule 1: IF x IS A2 OR y IS B3 THEN k is k2 K2 = 30
µ µ µ K1 K2 K3
A1 A2 A3 B1 B2 B3
1 1 1
AND
0.5 (min)
0.4
0.4
0 0 0
x1 X y1 Y k1 K
Rule 2: IF x IS A1 AND y IS B2 THEN k is k1 K1 = 1 0
b
µ µ µ
1 1 aggregation 1
0.8 0.8
0.4 0.4
0 0 0
k2 K k1 K k1 k2 K
µ µ
1 1
0.8 0.8
0.4 defuzzification 0.4
(weighted average)
0 0
k1 k2 K k1 k2 K
= = 23.33
Fig. 4 (a) Zero-order Sugeno-type fuzzy logic inference engine. (b) Aggregation and defuzzification
using trapezoidal and triangular membership functions, and applies a

set of inferential rules of the form:
IF x IS A AND y IS B THEN k is K
IF x IS A OR y IS B THEN k is K
IF x IS A THEN k IS K
The antecedent (input) part of the rule combines the fuzzyfied
inputs by AND (minimum) fuzzy operator or by OR (maximum)
fuzzy operator, or directly. The output of each rule is constant
(zero order), and then the consequents are represented by single-
ton membership functions.
To defuzzify the output, so a crisp value is available, the
weighted average (WA) method is applied (Fig 4b) (see Note 3):
WA =
åm ( x ) x (9)
åm ( x )
where μ(x) is the degree to which the inputs x belong to the appro-
priate fuzzy sets.
2.7 System Modeling Matlab environment and the DSP toolbox are used to code the
and Simulation signal processing algorithms, to run them in simulation mode, and
Environments then to export the source code as ANSI-C. The Data Acquisition
(DAQ) toolbox is used to collect data from the fingerprint sensor
and from the microphone.
2.7.1 ANN To model and simulate the FFBP-ANN the Matlab environment
plus the DSP toolbox is used to code the FFBP-ANN, to run it in
simulation mode, and then to export the source code as
ANSI-C. The Data Acquisition (DAQ) toolbox is used to collect
data from the fingerprint sensor and from the microphone.
2.7.2 FLE The FLE is a Sugeno-type fuzzy logic inferential paradigm. To

model and simulate it the Matlab environment plus the Fuzzy
Logic toolbox is used to code the FLE, to run it in simulation
mode, and then to export the source code as ANSI-C. The Data
Acquisition (DAQ) toolbox is used to collect data from the finger-
print sensor and from the microphone.
2.8 Digital Signal To implement the embedded system, a system-on-chip (SoC) proces-
Processor sor is used. This is an application-specific processor (ASP) optimized
for signal processing, with on-chip memory and peripherals.
A 32-bit floating-point DSP, architecturally optimized to
process data of different bit format (8-bit, 16-bit, 32-bit), has been
chosen. Due to its computing architecture peculiarity, it is able to
process efficiently the 8-bit data of the image pixels and the 16-bit
data of the voice samples. The 32-bit floating-point data format
ensures the required computing precision both for the image and
the speech processing.
218 Mario Malcangi
3 Methods
The biometric authentication system (Fig. 5) consists of three

processing layers, the feature-extraction layer, the matching layer,
and the fuzzy logic-based decision layer. The feature-extraction
layer uses signal processing-based algorithms for hard and soft fea-
ture extraction. The matching layer uses the hard computing
(DSP) to identify the hard features and the soft computing (ANN)
to identify the soft features. The decision layer uses the soft com-
puting (FLE) to fuse the identification scores with the soft
features.
3.1 Feature The feature extraction layer implements the biometric information
Extraction Layer capture (voiceprint and fingerprint) and the signal processing
algorithm (hard computing) methods for feature extraction.
3.1.1 Voiceprint The voiceprint is captured by a voice-type microphone, which is

and Fingerprint Capture conditioned (linearized, amplified, and filtered), in the analog
domain and then 16-kHz sampled, and 16-bit quantized. The utter-
ance is optimally end-pointed [21] and segmented. The Hamming
window is applied to extract 10-ms frames from the speech-data
stream, using a 50 % overlap between adjacent frames to avoid data
loss at feature-extraction time.
hard-computing layer soft-computing layer

(digital signal processing) (neuro-fuzzy inference)
voiceprint
voiceprint score features
hard features authentication &
decisions fusion
speed & stress

soft features
soft
Artificial score Fuzzy authentication
Neural Logic score
Network Engine
soft features
total area &
mean intensity
fingerprint score
hard features authentication
fingerprint
Fig. 5 System architecture

The fingerprint image is captured by a 512 dot-per-inch (dpi)

solid-state fingerprint sensor. The image is available at the sensor
output as 64,512 pixels 8-bit quantized array (224 rows and 288
columns) image.
3.1.2 Feature Extraction Features (hard and soft) are extracted from the captured voiceprint
and fingerprint applying the hard computing algorithm. Data are
assembled in data structures useful to be processed at the pattern
matching layer.
Voiceprint hard features are structured as time-sequenced vec-
tors of data:
RMS(N)
ZCR(N)
AC(N)
CLPC(N)
where N is the vector length and each vector element is the
feature measurement at the window application time.
Fingerprint hard features are structured sequences of minutiae as:
MINUTAE(M)
where M is the vector length and each vector element is the
minutiae data.
From voiceprint the two soft features are measured as:
SPEED
STRESS
The two features are scalar data (one measurement executed
on the whole captured voiceprint data stream).
From the fingerprint two soft features are measured as:
TOTAL_AREA
MEAN_INTENSITY
The two features are scalar data (one measurement executed
on the whole captured fingerprint data array).
3.2 Matching Layer The matching layer implements the hard and soft computing scor-
ing systems. Each of these applies the appropriate algorithms for
scoring the input data referred to a set of template data belonging
to the persons to be identified. This layer runs in two different
modes, enrolling and identifying.
The enrolling mode is active when a new person is to be added
to the set of persons that are to be accepted. The identifying mode
is active when a person (allowed or not allowed) is accessing the
system furnishing its voiceprint and fingerprint. Both enrolling and
identification modes run the same pattern matching algorithms,
the first to store the reference templates, and the second to execute
the scoring.
220 Mario Malcangi
3.2.1 Enrollment Mode Voiceprint enrollment mode consists in storing as multiple tem-
plates the features measured each time the allowed person utters at
the microphone a specific (requested) vocal sequence. Fingerprint
enrollment mode consists in storing as multiple templates the fea-
tures measured each time the allowed person puts a finger
(requested) on the sensor.
3.2.2 Identifying Mode The voiceprint and fingerprint identifying mode runs the pattern
matching algorithms to score the input hard features related to the
stored templates. For each input (voiceprint and/or fingerprint) a
score is available at the matching layer output.
3.2.3 Soft Biometric The soft biometric scoring is executed running the FFBP-ANN.
Training and Scoring To do this the FFBP-ANN inputs the soft biometric features
and outputs the score according to how it learned about the soft
feature belonging to the authorized person. A training phase is
requested to embed the knowledge in the FFBP-ANN nodes
(neurons). This is run each time the soft features at the FFBP-
ANN inputs belong to an authorized person, explicitly during the
enrollment (training mode), and implicitly each time the person is
identified as authorized (evolving mode). Learning is executed
running the error back-propagation algorithm.
Back Propagating Networks (BPNs) are trained according to a
generalized least mean-square (LMS) algorithm:
w j ( k + 1) = w j ( k ) + h ( x t ( k ) - x ( k ) ) f j ( k ) (10)

The weights wj(k) are modified by the kth input activity pattern
fj(k), so that a new updated weight wj(k + 1) is available at k + 1th
input time. The modification is proportional to the difference
between the xt(k) target response and the current response x(k).
The constant η controls the learning rate and is in the range
0 < η < 1.
During the learning activity the difference between the target
activity and the current activity at the output layer is a learning
error indicator. To measure it, the total error E is measured as
K
1 N t
E =å å (xi (k ) - xiL ( k ))2 (11)
k =1 2 i =1
This is the total error measured at the N nodes of the output layer
L after K patterns of the training set have been applied. If E is mini-
mum, then weights at the end of the training are the best for the
L-layer BPN.
Applying the best trained weights set to the three-layer FFBP
(L = 3) enables this ANN to perform the scoring of the biometric
features. The ANN executes also the data fusion among the soft
biometric data from voiceprint and fingerprint.
3.3 Decision Layer The decision layer implements the data fusion and the decision
fusion information from the matching layer. This is executed by
the FLE that evaluates four kinds of data input:
●● Voiceprint score.
●● Fingerprint score.
●● Soft-biometric measurements.
●● Soft-biometric score.
Prior to a run, the FLE needs to be tuned for processing the
data inputs and producing the decision. Membership functions
and rules set have to be designed using the modeling and simula-
tion toolbox.
Input data are fuzzyfied using optimally tuned membership
functions (Figs. 6, 7, and 8a). Singleton membership functions are
applied for rule consequents (Fig. 8b).
The rules are tuned combining the fuzzyfied inputs as follows
(only the most significant are reported):
1. IF voiceprint_score IS high AND fingerprint_score IS high AND
soft_score IS high THEN authentication IS very high.
2. IF voiceprint_score IS medium AND fingerprint_score IS
medium AND soft_score IS high THEN authentication IS high.
3. IF voiceprint_score IS medium AND speech_speed IS high AND
soft_score IS medium THEN authentication IS high.
µ
very low low medium high very high
1
0
voiceprint score
µ
1
0
voice speed
µ
1
0
voice stress
Fig. 6 Membership functions to fuzzify inputs (voiceprint)

222 Mario Malcangi
µ
1
0
fingerprint score
µ
1
0
fingerprint speed
µ
1
0
fingerprint stress
Fig. 7 Membership functions to fuzzify inputs (fingerprint)
a
µ
low medium high
1
0
soft-features score
b
µ very low low medium high very high
1
0
authentication
Fig. 8 (a) Membership functions to fuzzify inputs (soft-features score) and (b) to defuzzify outputs
(authentication)
4. IF voiceprint_score IS low AND speech_speed IS high AND

speech_stress IS high THEN authentication IS average.
5. IF fingerprint_score IS medium AND fingerprint_total_area IS
high THEN authentication IS high.
6. IF finger_print_score IS low AND fingerprint_total_area IS
high AND fingerprint_mean_intensity IS high THEN authenti-
cation IS average.
7. IF voiceprint_score IS medium and fingerprint_score IS medium
AND soft_score IS low THEN authentication IS low.
8. IF voiceprint_score IS high and fingerprint_score IS low AND
soft_score IS low AND fingerprint_total_area_match IS low
AND fingerprint_mean_intensity_match IS low AND
speech_speed_match IS low AND speech_stress_match IS low
THEN authentication IS very low.
9. IF voiceprint_score IS low and fingerprint_score IS high AND
soft_score IS low AND fingerprint_total_area_match IS low
AND fingerprint_mean_intensity_match IS low AND
speech_speed_match IS low AND speech_stress_match IS low
THEN authentication IS very low.
Rules were derived from feature distribution. Each rule was
manually tuned using a fuzzy logic rule editor, the simulator, and
the knowledge of an expert:
●● Rule 1 is a reinforcement of voiceprint and fingerprint matchers.
●● Rule 2 combines a voiceprint matcher and a fingerprint matcher
when both scores are too close to the decision threshold.
●● Rules 3, 4, 5, and 6 act as recovery rules when the voiceprint
or fingerprint matchers generate a false rejection.
●● Rules 7, 8, and 9 protect against false acceptance.
For fine-tuning, many other rules can be generated to take
additional soft-biometric measurements into account. Using more
rules leads monotonically toward greater reliability in the
authentication process.
Trapezoidal and triangular membership functions are used to
process inputs. The inference rule set is then applied. The result of
all the rules is evaluated using the WA method, so the crisp output
value can be computed. Singleton membership functions were
used to defuzzify the final decision.
3.4 Performance Performance evaluation aims to measure the reliability of the

Evaluation implemented biometric identification method. Voiceprint and fin-
gerprint authentication were first implemented and tested sepa-
rately, and then jointly, and finally combined through the fuzzy
logic inference engine and the artificial neural network applied to
soft-biometric features. An “all-against-all” test strategy was
applied to obtain match and mismatch scores.
224 Mario Malcangi
Evaluation of joint voiceprint and fingerprint authentication

consists of taking as good the better of the two matches (OR).
Single-user authentication was performed, so this test had minimal
system requirements. The following results were produced:
●● Voiceprint alone: 90.5 % correctly accepted.
●● Fingerprint alone: 85.7 % correctly accepted.
●● Voiceprint OR fingerprint: 92.3 % correctly accepted.
●● Fuzzy logic decision fusion of voiceprint, fingerprint, and arti-
ficial neural network evaluated soft features: 95.8 % correctly
accepted.
The OR test confirmed that multi-biometrics can improve per-
formance compared to single-biometric authentication. System
performance can be significantly improved, while keeping com-
plexity to a minimum, using fuzzy logic as decision fusion and
reinforcing it with an artificial neural network applied to soft-
biometric features.
4 Notes
1. A dual microphone or an array of microphones needs to be

used when the biometric application is targeted to outdoor
applications and the person to be identified is close to other
persons. Beam forming can be implemented for noise reduc-
tion purpose.
2. The capacitive fingerprint sensors are implemented as a two-
dimensional or a mono-dimensional (strip) scanner. The per-
formance of the biometric system is not sensitive to this form
factor. The first is less computationally intensive and more
intuitive, but it is not optimal for system dimension reduction.
The second is computationally intensive because it implies that
the two-dimensional image has to be built by software and it is
less intuitive, but it is optimal for system dimension
reduction.
3. Weighted Average (WA) defuzzification method is a derivation
of the Centroid (Center of Gravity) method that fits well the
singleton membership function.
References
1. Jain AK, Pankanti S, Prabhakar et al (2004) Second generation biometrics: the ethical, legal
Biometrics: a grand challenge. Proc 17th inter- and social context. The International Library
national conference on pattern recognition of Ethics, Law and Technology, vol 11.
(ICPR), vol 2. pp 935–942 Springer, pp 49–79
2. Anil K, Jain AK (2012) Biometric recognition: 3. Baird SL (2002) Biometrics: security technol-
an overview. In: Mordini E, Tzovaras D (eds) ogy. The technology teacher 61(5):1, pp 8–22
4. Sujithra M, Padmavathi G (2012) Next genera- system- on-chip for real-time applications,
tion biometric security system: an approach for Banff, Canada, 6–7 July
mobile device security, Proc second interna- 15. Malcangi M (2004) Improving speech end-
tional conference on computational science, point detection using fuzzy logic-based meth-
engineering and information technology. odologies. Proc thirteenth Turkish symposium
ACM, New York, NY, USA, pp 377–381 on artificial intelligence and neural networks,
5. Corcoran P, Cucos A (2005) Techniques for Izmir, Turkey, pp 10–11
securing multimedia content in consumer elec- 16. Wahab A, Ng GS, Dickiyanto R (2005) Speaker
tronic appliances using biometric signatures. authentication system using soft computing
IEEE Trans Consumer Elect 51:545–551 approaches. Neurocomputing 68:13–17
6. Jain AK, Ross A, Pankanti S (2006) Biometrics: 17. Kar B, Kartik B, Dutta PK (2006) Speech and
a tool for information security. IEEE Trans Inf face biometric for person authentication. Proc
Forensics Security 1(2):125–143 IEEE international conference on industrial
7. Jain AK, Nandakumar K, Lu X et al. (2004) technology, India, pp 391–396
Integrating faces, fingerprint, and soft biomet- 18. Jang-Hee Y, Jong-Gook K et al (2007) Design
ric traits for user recognition. Proc biometric of embedded multimodal biometric systems.
authentication workshop, LNCS 3087, Prague, Proc of the int conf on signal image technolo-
pp 259–269 gies and internet based systems, SITIS
8. Hong A, Jain S, Pankanti S (1999) Can multi- 2007:1058–1062
biometrics improve performance? Proc 19. Jain AK, Dass SC, Nandakumar K (2004) Soft
AutoID`99, Summit, NJ, pp 59–64 biometric traits for personal recognition sys-
9. Anil J, Lin H, Sharath P (2000) Biometric tems. Proc international conf on biometric
identification. Commun ACM 43(2):90–98 authentication, LNCS, vol 3072. Hong Kong,
10. Cappelli R, Maio D, Maltoni D et.al (2006) pp 731–738
Performance evaluation of fingerprint verifica- 20. Pak-Sum Hui H, Meng HM, Mak M (2007)
tion systems. IEEE Trans Pattern Anal Mach Adaptive weight estimation in multi-biometric
Intell 28:3–18 verification using fuzzy logic decision fusion.
11. Bishop C (1995) Neural networks for pattern Proc IEEE international conference on acous-
recognition. Oxford University Press, Oxford tic, speech, and signal processing ICASSP’07,
12. Ciota Z (2001) Improvement of speech pro- vol 1. pp 501–504
cessing using fuzzy logic approach. Proc of 21. Runkler TA (1997) Selection of appropriate
IFSA World congress and 20th NAFIPS Int defuzzification methods using application
Conf, 2 specific properties. IEEE Trans Fuzzy Sys 5:
13. Bosteels RTK, Kerre EE (2007) Fuzzy audio 72–79
similarity measures based on spectrum histo- 22. Abiyev RH, Altunkaya K (2007) “Neural net-
gram and fluctuation patterns. Proc Int Conf work based biometric personal identification”,
Multimedia and Ubiquitous Engineering Frontiers in the Convergence of Bioscience and
2007, Seoul, Korea, 27–28 April Information Technologies. pp 682–687
14. Malcangi M (2002) Soft-computing approach 23. O’Shaugnessy D (1987) Speech communica-
to fit a speech recognition system on a single- tion—human and machine. Addison-Wesley,
chip. Proc 2002 international workshop Reading, MA
Chapter 14
Using Neural Networks to Understand the Information

That Guides Behavior: A Case Study in Visual Navigation
Andrew Philippides, Paul Graham, Bart Baddeley,
and Philip Husbands
Abstract
To behave in a robust and adaptive way, animals must extract task-relevant sensory information efficiently.
One way to understand how they achieve this is to explore regularities within the information animals
perceive during natural behavior. In this chapter, we describe how we have used artificial neural networks
(ANNs) to explore efficiencies in vision and memory that might underpin visually guided route naviga-
tion in complex worlds. Specifically, we use three types of neural network to learn the regularities within
a series of views encountered during a single route traversal (the training route), in such a way that the
networks output the familiarity of novel views presented to them. The problem of navigation is then
reframed in terms of a search for familiar views, that is, views similar to those associated with the route.
This approach has two major benefits. First, the ANN provides a compact holistic representation of the
data and is thus an efficient way to encode a large set of views. Second, as we do not store the training
views, we are not limited in the number of training views we use and the agent does not need to decide
which views to learn.
Key words Visual navigation, Computational neuroscience, Machine learning, Boosting, Restricted
Boltzmann machine, Infomax
1 Introduction
All animals are faced with the challenge of extracting useful infor-
mation from the environment and using that information to shape
their behavior in an adaptive way. A major goal for neuroscience is
to understand how neural circuits are organized to achieve that
process. This goal depends on understanding the information pro-
vided by the environment as well as the information required for a
particular behavior. For many animals, the primary sensory channel
is vision and scientists have long sought to understand how vision
drives behavior and underpins perception by asking questions such
227
228 Andrew Philippides et al.
as “What does the frog’s eye tell the frog’s brain?,”1 or “How do
humans recognize their grandmother?”2
If we take the example of a very simple visually guided behav-
ior, we can begin to understand the style of information process-
ing that we see in biological systems. Many animals perform
phototaxis, where they try and move in the direction of the bright-
est light; for instance to gain warmth or rise to the surface of the
sea. For some microorganisms, such as zooplankton, we know in
detail the visual circuits that produce this behavior. They have two
photoreceptors which detect light from one side of the body only.
These photoreceptors are directly connected to motile hair cells
[1] which beat rhythmically to propel the organism through the
world. Light on one side of the organism stimulates one photore-
ceptor, which inhibits the movement of the hair cells on the same
side of the organism. The undamped beating from hairs on the
“dark” side of the body will ensure a turn towards the light.
Evolution tends not to produce overelaborate solutions; we see
that the zooplankton has a minimal yet perfectly efficient sensory
circuit for the task in hand.
For more complex visually guided behaviors, however, it is
nontrivial to understand what an efficient sensory system would
look like and we are faced with the task of being objective about
animals with visual systems that are very different from ours, who
have different perspectives on the world [2, 3] and different
behavioral requirements. As it is objective in terms of the way it
treats sensory data, a good first step in this endeavor is to try to
explore regularities in the sensory information perceived by animals
during their natural behavior.
One available method, which can be used to objectively explore
sensory data, is to use artificial neural networks (ANNs), or machine
learning more broadly, to explore the information available in a
raw sensory array. Such uses of ANNs provide a mapping from the
sensory environment to behavior. Essentially ANNs can be used to
interrogate complex data and find the regularities and affordances
that biological agents might well be tuned to (rather than being
used as explicit models of animal brains). Here, we present case
studies describing how we have used ANNs to understand how
animals might use visual information for navigation.
1
Here Lettvin and colleagues found that parts of the frog’s visual system are
tuned to detect small moving objects. Thus the eye is providing information
to the frog about where to direct the tongue in the hope of catching a fly.
McCulloch and Pitts, authors of this paper, are commonly held up as pioneers
of neural network research. J. Y. Lettvin, H. R. Maturana, W. S. McCulloch,
and W. H. Pitts, “What the Frog’s Eye Tells the Frog’s Brain,” Proc. IRE 47
(1959) 1940–1951.
2
Lettvin again came up with the concept of a sparse code in the brain where
small numbers of individual neurons may be highly selective to concepts such
as one’s grandmother. See Quian-Quiroga, R., Fried, I., and Koch, C. (2013)
Brain Cells for Grandmother. Scientific American, Feb.
Visual Navigation 229
Navigation is an essential task for most animals [4] and indeed

some artificial autonomous systems. One of the champions of this
behavior is the ant whose foragers spend much of their working life
travelling huge distances searching for food and then accurately
and efficiently bringing that food back to their nest along idiosyn-
cratic habitual routes [5]. Ants possess a number of mechanisms
for orientation, including for some species social cues provided by
pheromones [6]. However, for many ants the principal source of
information for navigation comes from the visual scenes that they
experience when travelling along their familiar routes [7].
Behavioral experiments have shown us the efficient way that
ants use vision for navigation which depends on a fundamental
property of the visual world. In a complex world, two photographs
taken with the same camera can only be identical when the camera
location and orientation are matched. However, Zeil et al. [8]
showed that the difference between images taken from different
nearby locations is minimized when the orientations match. This is
also true for natural visual systems. Thus, if a view of the world
from a forward-facing camera is memorized when travelling along
a route, that memory can be used later to recover the original
direction of travel from nearby locations, by finding the best match
with the currently perceived view [8, 9]. In behavioral terms for
the ant this means scanning the world and finding the direction
that provides the best match with their memory [10]. Using this
simple strategy the required knowledge for an experienced ant for-
ager is equivalent to the views required to set the appropriate direc-
tions at all locations along their foraging routes.
What kind of challenge does this present to an ant? The views
experienced by ants are of low resolution (Fig. 1) with an almost
panoramic scene being encoded with the equivalent of 1,000 pix-
els. What’s more their visual systems, unlike those of flying insects,
are not particularly fast, so the ant may have an effective “frame
rate” of roughly ten views per second. However, we know that ants
easily learn routes 10s of metres in length and, given their walking
speed, we can estimate that along a single foraging route ants will
experience views of the world totalling millions of pixels.
Theoretically, if ants had perfect memories they could store the raw
pixel values for every view experienced along a route and use those
perfect “photographic” memories for navigation. However, this
seems unlikely given a total brain size of 500,000 neurons. Surely
evolution has been able to come up with a more efficient system
both in terms of efficiencies in the way a visual system encodes a
view and also in the way visual information is stored. Here we
describe how we have used ANNs to explore efficiencies in vision
and memory that might underpin visually guided navigation in
complex worlds.
Fig. 1 (a) The world as seen by a human eye. The human field of view is quite large, almost 120° wide.
However, we only have high-resolution vision at the center of our visual field, which we make up for by moving
our eyes. (b) An ant’s eye view of the world. Ants have a horizontal field of view spanning almost 360°, but with
homogenous low resolution. Notice how you cannot recognize the house in the ant’s eye image. This picture
shows a vertical angle of 45°, but ants do see at increased elevations, because they extract celestial compass
information
2 Our Approach
Biological background: Ants can learn the information to visually

guide a route after traversing it once [11]. Subsequently, they
habitually travel within a corridor defined by the views experienced
along the route. These stored views define the directions in which
to travel [12, 13] for all locations along the route. Perhaps surpris-
ingly, ants do not seem to learn discrete waypoints representing
key locations along a route [9]. Further, ants accomplish this task
with a low-resolution visual system (Fig. 1) which is not well suited
to object recognition but provides ants with a coarse visual encod-
ing representing the broad layout of terrestrial objects contrasted
against the sky [14, 15].
Problem from a machine learning perspective: Given a single route
traversal (to gather training images), we need to learn a compact
representation which encodes the set of training images such that
they can subsequently be used to navigate the route. Specifically,
the network which has been used to encode the views should be
able to be used to set a direction of travel given the c urrent visual
input from locations along or close to the original route.
Insight: We can use the views experienced in the first route traversal
as training data for an ANN designed to perform familiarity judge-
ments. After training the ANN will be able to output a familiarity
score for any novel view presented to it.
The fact that ants are moving, and therefore facing, in the cor-
rect direction most of the time during the first route traversal
means that these training views implicitly define the movement
directions required to stay on the route. Thus, it is sufficient to
learn the views as they are experienced. After the initial route
learning, the problem of navigation is reframed as a search for
views similar to those associated with the route. If a novel view is
familiar, it is likely that a similar view has been experienced before
and so the ant is likely to be close to the route and heading in an
appropriate direction. So, by intermittently scanning the environ-
ment and moving in the direction that gives the most familiar view
an ant can reliably retrace its route.
To learn a route we thus train an ANN to learn the regularities
within a series of views in such a way that it can make a familiarity
judgement about any new view presented to it. This approach has
two major benefits. First, an ANN is an efficient way to encode a
series of views. We do not attempt to store every view experienced
along the route, but instead use them to train the ANN. Second,
and as a corollary, because we are not storing all the views used
for training, we are not limited in the amount of training views we
can use and so the agent does not need to decide when or which
views to learn.
There are many different possibilities for training ANNs for
this task. The only prerequisite for our approach is that the ANN
should provide a measure of the familiarity of a presented view. In
our case we also wanted biological plausibility, in the sense that the
vision, memory size, and learning behavior should be plausibly
achieved by an ant. These requirements have driven our research
trajectory.
3 Approach 1: ANN as a Classifier
Our first attempt at modelling route navigation with an ANN [16]

used training data to establish a classifier which could determine
whether a given view is an on-route view (i.e., facing the correct
direction) or an off-route view (i.e., facing in a non-route direc-
tion). When novel views were presented to the classifier the associ-
ated confidence in the classification could then be used as a proxy
for the familiarity of the presented view.
In this approach, a robot with a panoramic camera is first
moved along a predefined route to gather the training images
(every 4 cm) along the route. In order to train a classifier it is nec-
essary to generate positive and negative training examples of the
input to be classified. In our case this means collecting views that
are part of the route and views that are not part of the route. The
positive examples are simply the forward-facing views experienced
along the route. The negative views consisted of views from the
route taken facing to the left and right of the direction of move-
ment at an angle of 45° relative to the route heading (Fig. 2a).
Classifying views is a difficult task because of the high dimen-
sionality of the input if one adopts a pixel-by-pixel representation. In
order to make learning more tractable one can project this high-
dimensional space into a lower dimensional space by passing views
through a set of visual filters. Following Viola and Jones [17] we
chose to construct a classifier using randomly generated filters com-
posed of Haar-like features. Haar-like features act like edge detectors
or crude approximations to Gabor filters and are maximally activated
at high-contrast boundaries in the image (Fig. 2b). The set of out-
puts from these feature detectors form the basis of our image repre-
sentation and are used to train a classifier via the Adaboost learning
algorithm, a commonly used variant of boosting [18].
Boosting is a supervised learning technique for constructing a
strong classifier from a set of weak classifiers given a training set of
labelled positive and negative examples. A weak classifier is one
that performs only slightly better than chance. Conversely, a strong
classifier is one that performs arbitrarily well. A strong classifier is
constructed from a linear weighted combination of the outputs of
weak classifiers. In our context, each visual filter is a candidate weak
classifier, hj(x), and consists of a Haar feature fj, a threshold θj, and
a parity pj that determines whether the output should be greater or
less than the threshold
ì1 if p j f j ( x ) < p j q j
hj (x ) = í (1)
î0 else
This process is illustrated in Fig. 2c–d We initially generate a large
pool (5,000) of such filters of random types and sizes and at ran-
dom positions in the visual field. We then use Adaboost to select
and combine a prespecified number, T, of these weak classifiers into
a strong classifier. By specifying T, we thus control the complexity
of the classifier. As an aside we highlight an interesting outcome of
the boosting process. Because Adaboost starts with a large pool of
feature detectors, and then performs feature selection to pick out
and use only those features that are most useful for the current
classification problem, we can interrogate sets of filters to see if
they relate to biologically salient features or portions of the world.
The Adaboost algorithm works as follows. At each iteration,
the training data xi (m on- and l off-route views) are weighted
according to a distribution that indicate the current importance of
each example in the dataset, with initial weights w0,1 set as 1/(2m)
or 1/(2l) for on- and off-route views, respectively. The weak clas-
sifiers are evaluated according to the weighted error
a Negative b
-ve view
-ve
Positive
+ve view
+ve
Negative
view
-ve
-ve
c Decision Boundary
Negative Examples
0.4
Positive Examples
0.3
Frequency
0.2
0.1
0
0 50 100 150 200 250 300 350
Feature Detector Output
d
*
if > 200 THEN h(x) = 1
*
if < 200 THEN h(x) = 0
e f
1m
1m
Fig. 2 Using an ANN as a classifier of on- and off-route views. (a) Training images are collected in three directions
every 4 cm along the training route. A forward-facing image is collected as an example of an on-route view and two
off-route views are also collected at ±45° relative to the route heading. (b) Examples of the four different classes of
Haar-like feature detectors, from left to right, a single block, edge-detector, Gabor-like edge detector, and diagonal
edge detector. Each feature detector can vary in size, shape, and position. As images are panoramic, feature detec-
tors wrap-around if they extend beyond the left or right edge of the image. White is a weight of 1, grey 0, and black
−1. (c–d) Constructing a weak classifier using a single Haar-like feature. Each training image is convolved with the
feature detector to determine an optimal threshold which determines the output for the unseen data. (c): Distribution
of feature detector outputs for the two classes for a Haar feature f. A threshold, θ, is determined that optimally sepa-
rates the distributions (vertical bar ). (d) The final weak classifier is defined by a feature detector, a threshold, and a
parity. For the example shown, the output of the weak classifier, h(x), for the image, x, is 1 if the feature detector
output exceeds a threshold of 200, or 0 otherwise. (e–f) Learning a circuit of the gantry workspace. The dark line
represents the training route. (e) The gantry workspace as viewed from above. (f) Performance of the algorithm from
ten different start positions (dashed lines ) using a boosted classifier with 50 features
m +l
et , j = åwt ,i h j ( xi ) - y i (2)
i =1
where yi is 1 for on-route and 0 for off-route views. If the mini-
mum error across the classifiers at iteration t, et, is 0, the algorithm
stops. Otherwise, the classifier associated with the lowest error is
denoted as ht, added to the strong classifier and removed from the
pool of features. The weights associated with the training data are
then updated according to
1-ci
æ e ö
wt ,i +1 = wt ,i ç t ÷ (3)
è 1 + et ø
where ci = 0 if xi is classified correctly and 1 if not, with weights
normalized after each iteration, so they sum to 1. Essentially, this
process increases the weights of incorrectly classified views relative
to correctly classified views, thereby encouraging the next weak
classifier to focus on examples that were incorrectly classified at
the last iteration. Weak classifiers are added until the overall clas-
sification performance exceeds a threshold or the maximum num-
ber of weak classifiers, T, is reached. The final classification is then
æT ö
H ( x ) = sign ç åat ht (x ) ÷ (4)
è t =1 ø
with an associated confidence value of

T
C (x ) = åa h ( x )
t t (5)
t =1

which is simply the degree to which the sum of the combined weak
classifiers differs from zero, prior to the sign being taken. This con-
fidence value is key to our approach as it is used to determine the
most likely direction of travel in subsequent navigation.
To navigate a route the classifier is fed multiple views in different
directions from a given position. By weighting each of the viewing
directions that produce positive classifications by their associated
confidence values, a movement direction is determined which is
most likely to keep the agent on the learned route. To demonstrate
the efficacy of this approach we implemented it on a large Cartesian
gantry robot equipped with a panoramic camera moving through
a 3 m × 2 m environment (Fig. 2e). During a training run a series
of views are collected for on- and off-route directions (Fig. 2e).
These views are used to train our classifier following the proce-
dures outlined above. During route recapitulation the camera is
positioned at the start of the route facing in the correct direction.
From this position views are sampled in a range of directions from
−60° to +60° in steps of 5° relative to the current heading. Views
are passed into the classifier and all of the viewing directions that
produce a positive classification contribute to a weighted a verage
which determines the direction of travel and a 5 cm step is made in
this direction. The process is then iterated.
As can be seen (Fig. 2f), the approach works well and robust
route navigation is achievable. As is typical of ants, the results show
that the agent navigates along a corridor defined by the training
path but influenced by noise and the environment. It was also
noted that the routes “smoothed out” some of the kinks that were
present in the original training run. While results were dependent
on the number of classifiers used (more classifiers resulted in better
performance) near-perfect results were achieved with 50 classifiers,
while very robust results could still be achieved with only 20 clas-
sifiers. We have thus shown that it is possible to learn a nontrivial
route through an environment using a simple view classification
strategy based on positive and negative views collected during a
single episode of learning. By using an ANN as a classifier, we
reframed the problem of route navigation in terms of a search for
familiar views. The benefits of the ANN are that it both provides a
compact, holistic representation of the information required for
route navigation but also a measure of the expected uncertainty of
the classification.
We have thus demonstrated the principle that ANNs can be used
to learn a holistic representation of the views that determine a route.
Secondly, we have shown that a coarse visual system CAN encode a
complex world because it is used to drive a specific behavior through
familiarity recall rather than object or place recognition.
4 Approach 2: Training an ANN Without Negative Examples
One issue with the solution presented above is the biological-

plausibility of the training as it requires both on- and off-route
views. Additionally, an iterative Adaboost algorithm does not seem
viable for implementation in the brain.
For our second approach we used the same general method of
employing a training route to gather data but in a simulated world
which mimics a desert ant habitat and therefore has the inherent
richness of natural visual scenes (Fig. 3a). Additionally, we did not
want to use negative examples in training. If we have no negative
example to tell us we are wrong, an alternative approach is to learn
a compact approximation of the distribution of views encountered
during training which can subsequently be used to infer whether or
not a novel view is part of the original training set.
An efficient way to do this is to learn regularities in the views
experienced via sets of feature detectors which reduce the dimen-
sionality of the sensory data. One very successful machine learning
approach to this kind of problem is the use of “autoencoder”
b
a
Hidden j
wij
Visible i
Training Recapitulation
Fig. 3 (a) The simulation environment. Top left : A typical simulated environment. Middle : High-resolution view
of the world from an ant’s perspective. View is panoramic, so forward is middle of the row and the view wraps
round from left to right. Bottom : Low-resolution representation of the view shown in middle panel. (b)
Architecture of a Restricted Boltzmann Machine. The visible nodes are associated with the input data and the
hidden nodes with the hidden variables of the model to be learnt. (c) Left: The training route taken through a
complex desert ant habitat. Upper right: A zoomed-in view of the familiarity landscape produced by the trained
network. Right: Ten recapitulated routes through the environment
networks [19] which are able to both learn a compact representa-

tion of the data distribution and, after training, auto-generate the
data. These processes can be thought of as encoding and decoding
the original data, effectively learning a generative model. In the
application of this approach to our ant navigation problem, the
way in which visual features are selected is not generated in a super-
vised fashion as in Approach 1, but is driven by statistical regulari-
ties within the set of views experienced as the ant follows a route
through the world.
A powerful implementation of “autoencoder” networks is as a
Restricted Boltzmann Machine (RBM) [20, 21]. Hence we
attempt to learn a compact representation of the distribution of
views experienced along the route by training an RBM on the set
of training views. Once the RBM has been trained we can use it to
infer the probability that a novel view was part of the training set.
Behavior is now driven by scanning the world and searching for the
viewing direction with the highest such probability, i.e., the one
most likely to be part of the training route.
The basic RBM architecture is shown in Fig. 3b. Nodes in the
network are stochastic binary neuron-like elements. The visible
nodes are associated with the visible variables (in our case the pixel
values from on-route views) and the hidden nodes are associated
with the hidden variables (or “causes”) inherent in the data (in our
case the higher level features). The layers are joined by symmetri-
cally weighted connections in the manner shown in the figure, but
there are no visible-visible or hidden-hidden connections. This latter
constraint is what makes this class of network “restricted” in relation
to the standard fully connected Boltzmann Machine [22, 23] and
greatly simplifies learning and inference [21]. In general, RBMs
can be stacked to create several hidden layers where the outputs of
one layer act as the data for training the next layer (producing so-
called deep networks) [19]. In our ant problem application a single
hidden layer sufficed—we used an RBM with 1,360 visible units
and 500 hidden units.
The stochastic nature of the RBM network nodes means that
it is natural to import methods and language from statistical phys-
ics in describing their operation. Hence we talk about a joint con-
figuration (v,h) of the visible and hidden units having an “energy”
given by the function E(v,h) shown in Eq. 6
E ( v,h ) = - å av i i - å b j h j - åvi h j (6)

i Îvisible j Îhidden i, j

where vi and hj are the binary states of the visible unit i and hidden
unit j respectively, ai, and bj are their biases and wij is the weight
between them. Through this energy function it is possible to
express the probability of a given configuration by
- E ( v ,h )
e
P (v,h ) = (7)
Z
where Z is a normalization constant. Thus, the probability assigned
to a single data point, v, is found by summing over all possible hid-
den vectors as shown in Eq. 8
åe
- E ( v ,h )
P (v ) = åP (v,h ) = h
(8)
h Z
Following a maximum likelihood principle, training an RBM con-
sists of increasing the probability that the network assigns to the
observed data by adjustment of the weights and biases. This is
achieved by decreasing the energy of the observed data and increas-
ing the energy of unobserved configurations using some kind of
gradient descent algorithm [20, 21]. There are many possibilities,
but in the current work we employ an efficient approach called Fast
Persistent Contrastive Divergence; full details of the training pro-
cedure can be found in Hinton [21, 24] and its application to the
ant problem in Baddeley et al. [25].
Having trained an RBM, the probability of a visible datum (in
our case the visual scene) is found by summing over all possible
configurations of the hidden units as described by Eq. 8.
Unfortunately this computation is intractable for large h. However,
we can calculate the probability of a visible datum up to a normal-
ization constant by using an alternative, tractable, formulation for
the probability of a datum expressed in terms of the free energy,
F(v), as described by Eq. 9:
F ( v ) = - log åe
- E ( v ,h )
(9)
h
- F (v )
which allows us to write P (v ) = e where Z′ is a normaliza-
Z¢
tion constant. For an RBM the free energy has an analytic form
given by Eq. 10:
F ( v ) = åai vi - å log 1 + e ( xj
) (10)
i j

where x j = b j + åwij vi . Since we wish to compare the relative
i
probabilities of different views, the normalization constants cancel
out. This allows us to directly compare the probabilities that each
of the views in a scan is part of the training route by simply calcu-
lating their free energies.
The routes we produced can be seen in Fig. 3c. As before, there
is a broad route corridor defined by the training data and the structure
of the environment. As with the classifier approach, this shows that an
ANN-based holistic memory can achieve robust route homing.
However, in this instance, no “negative views” were needed and so

the approach is more attractive in terms of biological plausibility. In
addition, a nice feature of this algorithm is that once trained, the net-
work can be queried to see the types of regularities that have been
encoded which again allows us to understand the visual filters that are
appropriate for the visual navigation.
Unfortunately, the training process for RBMs has a so-called
“burn-in” period before unbiased samples can be produced. This
results in a computation and time heavy algorithm. So while this
works satisfactorily without negative training views, it is not ideal
for our biological desideratum of one-trial learning.
5 Approach 3: Learning and Discarding Views
We have shown that it is possible to learn a compact holistic route

memory that gives a confidence that any novel view is part of that
route and that this can be done without off-route (i.e., negative
examples). However, so far we have used an ANN’s confidence in
its classification as a proxy for familiarity. In so doing, during ANN
training we have needed to have all the views available which seems
somewhat counterintuitive biologically; the agent would need an
almost perfect memory before then being able to acquire a com-
pact representation by training the network. We now turn to an
alternative network and training regime. Instead of storing all of
the views experienced on a training route, the views are used to
train a two-layered ANN to perform familiarity discrimination
using an Infomax learning rule [26]. That is, once trained, the
network takes a view as input and outputs a familiarity measure
based on the views that the network was trained with. The key dif-
ference to the previous ANN approaches is that each training view
is presented to the network once only and then discarded follow-
ing a single cycle of weight updating. Thus the memory load does
not scale with the length of route but remains constant.
The ANN architecture (Fig. 4a) consists of an input layer and
a novelty layer. The number of input units is equal to the dimen-
sionality of the input which in our case is the number of pixels in a
view of the world, N = 90 × 17 = 1,530. The number of novelty
units, M, is arbitrary and we typically used the same number of
novelty units as inputs, i.e., M = N = 1,530. We found that using as
few as 200 novelty units can still work well, but we did not explore
this aspect of the problem in any detail. The network is fully con-
nected by feedforward connections wij with the activation of the
i′th novelty unit being
M
ai = åwij x j (11)
j =1

a
Input
Layer
X1 XN
wij
Novelty
Y1 YM
Layer
2m
1m
Fig. 4 (a) The two-layered network used with the Infomax learning rule. Circles
represent units and arrows denote connections between the input units on the top
and the novelty units on the bottom. There is no output from the network as such
since the response of the network is a function of novelty unit activations. (b)
Navigating using a trained ANN to assess scene familiarity. Three separate routes
(thick grey lines) learned in an environment containing both small and large objects.
For each of the three routes, that consisted of between 700 and 980 views taken
every 1 cm, we show three recapitulations (thin black lines). During route reca-
pitulations the headings at each step were subject to normally distributed noise
where xj is the value of the j′th input pixel. Weights are initialized
randomly from a uniform distribution in the range [−0.5, 0.5] and
then normalized so that the mean of the weights feeding into each
novelty unit is 0 and the standard deviation 1. The Infomax
approach [27] decomposes each view into a fixed number of com-
ponents (determined by the number of hidden units in the net-
work) which remains constant, independent of the number of
views experienced. It does this by adjusting the weights so as to
maximize the information that the novelty units provide about the
input, by following the gradient of the mutual information. Since
two novelty units that are correlated carry the same information,
adjusting weights to maximize information will tend to decorrelate
the activities of the novelty units and the algorithm can thus be
used to extract independent components from the training data
[28]. The core update equation is
h é N
ù
Dwij =
N êwij - ( y i + hi ) åhk wkj ú (12)
ë k =1 û
where yi = tanh(hi) is the output of each novelty unit, and η = 0.01

is the learning rate. This performs gradient ascent using the natural
gradient [29] of the mutual information over the weights [28]
which avoids the computationally expensive calculation of the
inverse of the entire weight matrix.
Once trained, subsequent route navigation is achieved by the
agent sampling views at different headings from a given position.
Each view is fed into the ANN and the agent takes a 5 cm step in
the direction associated with the most familiar view where the
familiarity of a view, X, is given by the network output:
M
d ( X ) = å hi (13)
i =1
The network response could be viewed as an output layer but as it
is a function of the activations of the novelty units, we follow [26]
and do not represent it with another layer. The Infomax familiarity
measure is abstract and reflects whether an input is well described
in terms of the learned components that the hidden units repre-
sent. While with a limited amount of data the algorithm is unlikely
with a standard deviation of (15o). (c) Including learning walks prevents return
paths from overshooting the goal. Without a learning walk the simulated ant over-
shoots and carries on in the direction it was heading as it approached the nest
location. By including the views experienced during a learning walk the simulated
ant, instead of overshooting, gets repeatedly drawn back to the location of the
nest. Thick grey line: training path; thin black lines: recapitulations
to converge to a particularly good set of independent components,

it is enough that the components that are extracted provide a more
suitable decomposition of the training data than of an arbitrary
input. By decomposing the input in this way it is possible to compress
redundant data resulting in more efficient memory storage.
This process achieves robust route navigation, again displaying
a corridor within which route navigation is easily achieved (Fig. 4b).
The algorithm is able to learn multiple routes to a single goal and
can be made more robust by including a number of training runs.
Further, by including a set of views that face the goal from a num-
ber of different directions, inspired by the learning walk behavior
of ants [30], the same algorithm and ANN memory are able to
exhibit place search as well as route navigation (Fig. 4c).
6 Discussion and Future Directions
Our aim was to use artificial neural networks (ANNs) as a tool to

investigate the connection between the sensory environment and
behavior for the problem of visual navigation. It was hoped that
through this machine learning approach we might explore the com-
plex sensory array and find regularities and affordances that biologi-
cal agents might be tuned to. Over a series of implementations we
have shown how the visual scenes experienced during a single route
traversal can be used to train ANNs to learn a compact representa-
tion of the visual information needed to recapitulate that route.
Such recapitulation is driven by sampling the world in a range of
directions and assessing the familiarity of the associated views as
judged by the ANN; one then has to simply move in the most
familiar direction. As we have shown, this familiarity-based naviga-
tion works robustly in a number of environments.
A more general outcome that emerged from the use of ANNs
in this series of experiments was that they allowed us to objectively
assess the opportunities for efficient use of vision in a complex task
(visual navigation). Our first hope was that objective neural net-
work modelling would allow us to investigate what visual regulari-
ties are extracted by ANNs in complex natural worlds. In our first
two approaches we were able to see which visual features emerge as
useful descriptors of the world for navigation. We suggest this as a
new method by which biologists might try to understand the visual
preprocessing that is useful for a particular task.
A second major outcome has been a deeper understanding of
how memory systems might be efficiently organized for visual navi-
gation. Our algorithms are existence proofs that route memories
can be held within a single ANN, which is thus a “holistic” repre-
sentation of the route. This means that memory load does not scale
with the length of the training route and, for instance, multiple or
complex paths to a single goal can be stored together. From a
iological perspective it is pleasing that recent research suggests

b
that a particular brain region in insects (the mushroom body)
may be well suited to computations similar to our algorithms
(Barbara Webb, personal communication, June 10, 2014).
Another property of these “holistic” networks is that at no point
does the agent have a spatial memory in terms of knowing “where
it is” but rather it has a spatial memory for “what it should do.”
This has often been noted of insect navigators [15, 31].
Overall, we hope to have outlined in this chapter how neural
networks, and machine learning techniques more generally, are not
restricted to being used as specific models of brain areas, but can be
used to provide a deeper understanding of animal behavior.
References
1. Jékely G, Colombelli J, Hausen H et al (2008) ants (Cataglyphis velox). Behav Ecol 23(5):
Mechanism of phototaxis in marine zooplank- 944–954. doi: 10.1093/beheco/ars051
ton. Nature 456(7220):395–399 12. Wystrach A, Beugnon G, Cheng K (2012) Ants
2. von Uexküll J (1931) Der Organismus und die might use different view-matching strategies
Umwelt. In Driesch H, Woltereck H. (Eds.), on and off the route. J Exp Biol 215(1):44–55.
Das Lebensproblem im Lichte der modernen doi:10.1242/Jeb.059584
Forschung, Quelle und Meyer, Leipzig, pp 13. Collett M (2010) How desert ants use a visual
189–224 landmark for guidance along a habitual route.
3. Nagel T (1974) What is it like to be a bat? Proc Natl Acad Sci U S A 107(25):11638–
Philos Rev 83:435–450 11643. doi:10.1073/pnas.1001401107
4. Shettleworth SJ (2010) Clever animals and kill- 14. Graham P, Cheng K (2009) Ants use the pan-
joy explanations in comparative psychology. oramic skyline as a visual cue during navigation.
Trends Cogn Sci 14(11):477–481. Curr Biol 19(20):R935–R937
doi:10.1016/j.tics.2010.07.002 15. Wystrach A, Graham P (2012) What can we
5. Wehner R (2009) The architecture of the des- learn from studies of insect navigation? Anim
ert ant’s navigational toolkit (Hymenoptera: Behav 84(1):13–20. doi:10.1016/j.anbehav.
Formicidae). Myrmecological News 12:85–96 2012.04.017
6. Hölldobler B (1990) The ants. Harvard 16. Baddeley B, Graham P, Philippides A et al (2011)
University Press, Cambridge, MA Holistic visual encoding of ant-like routes: navi-
7. Collett TS, Graham P, Harris RA et al (2006) gation without waypoints. Adapt Behav 19(1):3–
Navigational memories in ants and bees: 15. doi:10.1177/1059712310395410
Memory retrieval when selecting and following 17. Viola P, Jones M (2001) Rapid object detec-
routes. Adv Study Behav 36:123–172. doi: tion using a boosted cascade of simple features.
10.1016/S0065-3454(06)36003-2 In: Computer vision and pattern recognition,
8. Zeil J, Hofmann MI, Chahl JS (2003) The 2001. CVPR 2001. Proceedings of the 2001
catchment areas of panoramic snapshots in out- IEEE Computer Society Conference on, 2001.
door scenes. J Opt Soc Am A Opt Image Sci IEEE, vol 511. pp I-511–I-518
Vis 20:450–469 18. Freund Y, Schapire R, Abe N (1999) A short
9. Philippides A, Baddeley B, Cheng K et al introduction to boosting. J Jpn Soc Artif Intell
(2011) How might ants use panoramic views 14(771–780):1612
for route navigation? J Exp Biol 214(3):445– 19. Hinton GE, Salakhutdinov RR (2006)
451. doi: 10.1242/Jeb.046755 Reducing the dimensionality of data with
10. Wystrach A, Philippides A, Aurejac A et al (2014) neural networks. Science 313(5786):
Visual scanning behaviours and their role in the 504–507
navigation of the Australian desert ant 20. Smolensky P (1986) Information processing in
Melophorus bagoti. J Comp Physiol A:1–12 dynamical systems: foundations of harmony
11. Mangan M, Webb B (2012) Spontaneous for- theory. In Rumelhart D, McClelland J, the
mation of multiple routes in individual desert PDP Research Group (Eds.). Parallel distributed
processing: Explorations in the microstructure familiarity discrimination and feature extrac-

of cognition. Vol. 1: Foundations. MIT Press, tion in a single network. Neural Comput
Cambridge, MA, pp 194–281 23(4):909–926
21. Hinton GE (2002) Training products of 27. Bell AJ, Sejnowski TJ (1995) An information-
experts by minimizing contrastive divergence. maximization approach to blind separation and
Neural Comput 14(8):1771–1800 blind deconvolution. Neural Comput 7(6):
22. Ackley DH, Hinton GE, Sejnowski TJ (1985) 1129–1159
A learning algorithm for Boltzmann machines. 28. Lee T-W, Girolami M, Sejnowski TJ (1999)
Cognit Sci 9(1):147–169 Independent component analysis using an
23. Hinton GE, Sejnowski TJ (1986) Learning and extended infomax algorithm for mixed sub-
relearning in Boltzmann machines. MIT Press, gaussian and supergaussian sources. Neural
Cambridge, MA, 1 (282–317):4.2 Comput 11(2):417–441
24. Hinton G (2010) A practical guide to training 29. Amari S-I, Cichocki A, Yang HH (1996) A new
restricted Boltzmann machines. Momentum learning algorithm for blind signal separation.
9(1):926 Adv Neural Inform Process Syst 8:757–763
25. Baddeley B, Graham P, Philippides A et al 30. Muller M, Wehner R (2010) Path integration
(2011) Models of visually guided routes in provides a scaffold for landmark learning in
ants: embodiment simplifies route acquisition. desert ants. Curr Biol 20(15):1368–1371.
Intelligent robotics and applications. Springer, doi:10.1016/j.cub.2010.06.035
New York, pp 75–84 31. Wehner R, Wehner S (1990) Insect naviga-
26. Lulham A, Bogacz R, Vogt S et al (2011) tion—use of maps or ariadnes thread. Ethol
An infomax algorithm can perform both Ecol Evol 2(1):27–48
Chapter 15
Jump Neural Network for Real-Time Prediction

of Glucose Concentration
Chiara Zecchin, Andrea Facchinetti, Giovanni Sparacino,
and Claudio Cobelli
Abstract
Prediction of the future value of a variable is of central importance in a wide variety of fields, including
economy and finance, meteorology, informatics, and, last but not least important, medicine. For example,
in the therapy of Type 1 Diabetes (T1D), in which, for patient safety, glucose concentration in the blood
should be maintained in a defined normoglycemic range, the ability to forecast glucose concentration in
the short-term (with a prediction horizon of around 30 min) might be sufficient to reduce the incidence
of hypoglycemic and hyperglycemic events. Neural Network (NN) approaches are suitable for prediction
purposes because of their ability to model nonlinear dynamics and handle in their inputs signals coming
from different domains. In this chapter we illustrate the design of a jump NN glucose prediction algorithm
that exploits past glucose concentration data, measured in real-time by a minimally invasive continuous
glucose monitoring (CGM) sensor, and information on ingested carbohydrates, supplied by the patient
himself or herself. The methodology is assessed by tuning the NN on data of ten T1D individuals and then
testing it on a dataset of ten different subjects. Results with a prediction horizon of 30 min show that
prediction of glucose concentration in T1D via NN is feasible and sufficiently accurate. The average time
anticipation obtained is compatible with the generation of preventive hypoglycemic and hyperglycemic
alerts and the improvement of artificial pancreas performance.
Key words Type 1 diabetes, Forecast, Continuous glucose monitoring, Nonlinear modeling, Time
series modeling
1 Introduction
Prediction of the future value of a variable is of central importance

and widely applicable within a variety of disciplines. For example,
as reported in ref. 1, economic forecasting is used in financial man-
agement, for setting fiscal policies and budgeting, and for business
planning. A variety of applications exist in both long- and short-
term weather forecasting, for example in the generation of civil
danger warnings, decision support in agriculture, planning of elec-
tricity demand, control and safety of air traffic. In computer s cience,
245
246 Chiara Zecchin et al.
forecasting the behavior of Internet networks is used to optimize

resources allocation and detect security attacks.
In medicine, prediction methods are largely used for tuning
therapies and for forecasting risks of development and progression
of diseases, see for example ref. [2] for a review of advances in
genetics for predicting the occurrence of some diseases. In subjects
affected by chronic pathologies, highly risky events could be fore-
casted by exploiting short-time prediction methods, with
20–30 min of anticipation, see [3] for an application in epilepsy for
impending seizure detection. In Type 1 Diabetes (T1D) manage-
ment, which is the field of application considered in the present
chapter, real-time short-term prediction of future glucose concen-
tration in the blood from its past history measured by minimally
invasive subcutaneous continuous glucose monitoring (CGM)
sensors [4, 5] could allow diabetic patients to administer treatment
and adjust their therapy on the basis of future, instead of current,
glucose concentration, permitting them to mitigate, and in some
cases avoid, critical events, see [6, 7] for methodological review
aspects, [8, 9] for clinical applications and [10, 11] for industrial
challenges. In addition, CGM and short-term prediction are key
inputs to the so-called artificial pancreas, a minimally invasive
pump device which subcutaneously administers insulin according
to a temporal profile determined in real-time by a sophisticated
closed-loop control algorithm, see for example refs. [12, 13].
In this chapter we present, in detail, the recent glucose predic-
tion algorithm of ref. 14, which is based on a jump NN model that
uses, as inputs, not only the information on glucose concentration
history measured by the CGM sensor but also the quantity of
ingested CHO provided by the patient in concomitance with the
meal. Parameters of the NN are first tuned on data of ten T1D
subjects, while the effectiveness of the prediction method is tested
on ten different T1D subjects monitored for 2–3 consecutive days
by a widely used commercial CGM sensor (Dexcom SEVEN®
PLUS, Dexcom Inc., San Diego, CA).
2 The Diabetes Disease and Its Therapy
Diabetes mellitus is characterized by dysfunctions in insulin secre-

tion and action: in T1D the pancreas is unable to produce insulin,
while in Type 2 diabetes derangements in insulin secretion and
action occur. As a consequence, the glucose concentration in the
blood, which in a healthy individual remains within the normal
euglycemic range of 70–180 mg/dl, often exceeds these limits,
with short- and long-term complications. Hypoglycemia (glycemia
below 70 mg/dl) can progress from measurable cognition impair-
ment to aberrant behavior, seizure and coma [15]. Hyperglycemia
(glycemia above 180 mg/dl), if left untreated, can become severe
Glucose Prediction by Jump NN 247
and lead to serious complications requiring emergency care, such

as diabetic coma. Moreover, prolonged hyperglycemia predisposes,
in the long term, to invalidating pathologies, as neuropathy,
nephropathy, retinopathy, and diabetic foot ulcers [16].
Conventional T1D therapy aims at maintaining euglycemia by
adjusting diet, physical activity, and insulin injections. The therapy
is normally tuned on the basis of 3–4 daily fingerstick self-
monitoring blood glucose (SMBG) measurements, obtained by
the patient through portable lancing devices [17]. The recent
development of portable minimally invasive CGM sensors, able to
measure glucose concentration every 5–10 min for up to 7–10
days, allows the tracking of glucose dynamics much more effec-
tively than via SMBG. It is today largely accepted in clinical research
that CGM sensors permit the improvement of diabetes manage-
ment [18–20], both by suggesting the refinement of the patient’s
individual therapy on the basis of the retrospective (Holter-like)
assessment of glycemic recordings and in real-time, by alerting the
patient when hypoglycemic and hyperglycemic thresholds are
exceeded. However, taking therapeutic decisions on the basis of
the current glycemia does not allow the patient to avoid imminent
critical events.
To help readers of this book to better understand this issue, in
Fig. 1 the time-course of glucose concentration (black dots linearly
interpolated) measured during the day of a T1D subject by a CGM
sensor, together with information on timing and quantity of
ingested carbohydrates (CHO, blue stems) and timing and quan-
tity of injected insulin bolus (green stems) are shown. Hypoglycemic
and hyperglycemic thresholds are also reported (thin horizontal
lines). At the end of the first night glucose concentration was in the
euglycemic range, but fell below 70 mg/dl around time 07:30.
The subject had breakfast around 08:00 but did not inject any
insulin bolus in concomitance with his meal. During the morning
300
CGM[mg/dl]
insulin[U]
250 CHO[g]
200
hyperglycemic threshold
CGM[mg\dl]
150
100
hypoglycemic threshold
70g CHO 70g CHO
50
40g CHO
5U insulin 10g CHO 5U insulin
3U insulin
0
Tue 06:00 Tue 08:00 Tue 10:00 Tue 12:00 Tue 14:00 Tue 16:00 Tue 18:00 Tue 20:00 Tue 22:00 Wed 00:00 Wed 02:00 Wed 04:00
time[Day HH:MM]
Fig. 1 Representative CGM signal (black dots linearly interpolated) measured by the Dexcom SEVEN PLUS
device and information on insulin doses (green stems) and meals (blue stems) of a T1D subject
glucose concentration reached hyperglycemic values and the

subject injected a correction bolus of insulin around time 10:00
and reentered the euglycemic range around time 12:00. At time
13:00 and 19:00 the subject ate and injected insulin to counterbal-
ance the effects of CHO. Notably, around time 17:00 the CGM
signal fell in the hypoglycemic range and the subject promptly
ingested 10 g of sugar to increase the glucose concentration and
reenter the safe range. After dinner, around time 20:00, the subject
experienced another hyperglycemia and reentered the safe range
only after time 01:00. Some of (or even all of) these critical events
could have been avoided, in theory, if the alerts were generated
about 20 min before the crossing of the 70 mg/dl threshold, e.g.,
on the basis of a predicted profile of future glucose concentration.
In view of the availability of CGM sensors from the beginning of
the twenty-first century this possibility stimulated research and
development of glucose prediction methods [4, 7]. However,
forecasting glucose concentration in a certain prediction horizon
(PH) is a challenging topic. Indeed, glycemia is influenced by
many (often not measurable/quantifiable) factors (e.g., CHO
ingestion, physical activity, administration of drugs including
insulin, stress, and emotions) and interindividual and intraindi-
vidual variability is high. Furthermore, information relative to sev-
eral key variables that influence glucose dynamics is not directly/
easily usable. For instance, glucose concentration is measured by
the CGM sensor and is returned as a time series, thus it is directly
available and easy to embed in formal mathematical models. On
the other hand, information on timing and CHO content of meals
and on timing and dose of injected insulin is impulsive and should
be adequately preprocessed, before being used, to generate signals
more informative of meal and insulin effects on glucose concen-
tration. Finally, information on factors such as stress or emotions
is hard to quantify and highly subjective. For these reasons, the
majority of published glucose prediction methods rely solely on
the use of the CGM signal as input.
3 Glucose Prediction: A Brief State of the Art
Popular algorithms, which exploit CGM information only, include

autoregressive (AR) and AR with moving average (ARMA) mod-
els. Some examples are [21–24]. Some attempts to exploit infor-
mation on CHO and insulin therapy by ARX and ARMAX models
have been recently proposed, see for example ref. [25–28].
However, it is natural to expect that glucose prediction performed
exclusively on the basis of CGM data, possibly incorporating addi-
tional information in simple linear models, cannot be as effective as
if additional input signals were used in nonlinear models.
Neural networks (NN), thanks to their ability to learn the

behavior of empirical nonlinear models and to utilize heteroge-
neous signals among their inputs, are valuable models for forecast-
ing glycemia using collateral information, in addition to the CGM
signal. In the literature, only a few papers report investigations of
the use of NN algorithms for glucose prediction. In ref. 29 the
authors proposed a feed-forward NN whose inputs were previous
CGM samples and the current time instant. In ref. 30 Pappada and
colleagues developed a feed-forward NN incorporating, in addi-
tion to CGM data, other inputs such as SMBG readings, informa-
tion on insulin and CHO, information on hypoglycemic and
hyperglycemic symptoms, lifestyle, activity, and emotions. In ref.
31 Daskalaki and colleagues compared an ARX and a NN model
exploiting, respectively, CGM and insulin and CGM, insulin and
CHO information. In ref. 32 our research group proposed a model
based on the parallel of a first order polynomial algorithm and a
feed-forward NN, using information on CGM and announcement
of future ingested CHO. In ref. 14 we demonstrated that a simpler
architecture, without any announcement of future ingestion of
CHO, gives results comparable to those of ref. 32.
4 A Jump Neural Network Methodology for Glucose Prediction
A jump NN is a feed-forward NN with inputs connected not only

to the first hidden layer but also to the output layer. According to
ref. 33, the jump NN structure is particularly suitable, and better
than a simple feed-forward NN, for fitting and predicting time
series characterized by the presence of both linear and nonlinear
dynamics, as in the case of glucose signals, where we assume that
each input is able to influence future blood glucose concentration
with both linear and nonlinear effects. Indeed, hidden neurons,
with their nonlinear activation functions, model the nonlinear rela-
tionship between inputs and targets, while the output neurons,
with their linear activation functions, learn the linear relationship
between inputs and targets.
To explain the adopted methodology, Fig. 2 shows the archi-
tecture of the proposed jump NN. Note that, in our implementa-
tion, data have a sampling period Ts of 5 min and the PH is 30 min,
corresponding to N = 6 steps ahead prediction.
4.1 Inputs Selection Inputs were selected through a multiple consecutive steps proce-
and Preprocessing dure, using data drawn from the training set.
1. First, a pool of possible candidate input signals was chosen,
based on a priori physiological knowledge of the glucose insulin
system and of diabetes. Going into details, the history of the
glucose concentration signal, measured by the CGM sensor,
Fig. 2 Block scheme of the proposed jump NN model for glucose prediction. z− k indicates the k-steps back-
ward shift operator
was chosen since glucose concentration is a slow-varying signal,

thus its future values are correlated, in the short term, with the
past time-course of the signal. Moreover, information on quan-
tity and timing of ingested CHO, which is assumed to be pro-
vided by the patient himself/herself, was used. Ingestion of
sugar is known to increase glucose concentration; however,
while the ingestion of a meal can recall an impulsive process,
CHO effects on glycemia are known to be smoother in both
appearance and disappearance. For this reason, the physiologi-
cal model of oral glucose absorption proposed in ref. 34, com-
pleted with the population parameters obtained in ref. 35, was
used to obtain the glucose rate of appearance (RaG) in the
blood, a signal that mimics the velocity of glucose absorption in
the blood stream after the ingestion of a CHO load.
2. As second step, various signals, related to CGM history and to
RaG were generated. The glucose concentration time derivative
was computed from the CGM signal, using a Bayesian smoothing
approach [36] to deal with ill-conditioning which tends to amplify
measurement noise. RaG first order difference was computed,
with time steps of 5, 15, and 30 min, and RaG cumulative sum in
the preceding 30, 60, 90, and 120 min was also computed.
3. Afterwards, the cross-correlation between target glucose con-
centration and the signals generated at step 2 was computed,
for various time shifts. This step allows one to obtain two major
results: the signals with the highest correlation with the target
are selected, which are likely to be more useful to predict the
target itself; time delays that should be applied to these signals
are also selected.
4. Finally, k-fold cross-validation was used to pick out the best

subset of inputs, among those selected with the cross-
correlation analysis. In this step, all the possible input combi-
nations are considered and, for each one, a NN is optimized.
In particular, in k-fold cross-validation the training set is
divided into k subsets, each network is trained on k − 1 subsets
and tested on the remaining subset and the procedure is
repeated k times, leaving out each time a different subset for
testing the models. The performance of each NN is the aver-
age obtained in the k experiments.
The above four steps provided the data necessary to select four
input signals. Two of them are related with the CGM signal and
are the current value of glucose concentration, measured by the
CGM sensor, indicated as y(n) in Fig. 2 and its current time deriva-
tive, which will be indicated as Δy(n) in the rest of the chapter. The
other two inputs are relative to information on ingested CHO and
are the current value of RaG and its average trend in the last 30 min,
computed as
1 1
N
( 1 − z −N ) RaG (n ) = ( RaG (n ) − RaG (n − N ) )
N
(1)

with z− k indicating the k time steps backward shift operator, n cur-
rent time instant, and N equal to six steps.
Several preprocessing techniques are commonly applied before
the data are used for training the NN to accelerate convergence
and to ease the problem to be learned [37]. An essential operation
is a rough scaling of the data between the lower and upper bounds
of the neuron transfer function. We normalized inputs and output
so that they had zero mean and standard deviation equal to 1. This
step guarantees that, at the beginning of the training procedure, all
the signals have, potentially, the same importance and they all
belong to the approximately linear range of the tangent sigmoidal
activation function. Data normalization also limits the extent to
which larger numbers may dominate smaller ones and avoids pre-
mature saturation of hidden nodes, which would degrade the
learning process.
4.2 Mathematical Let us define: Nin number of input signals and Nhn number of hid-
Representation den neurons. The jump NN predicts a signal that may be expressed,
of the Jump NN Model in a vector matrix form, as
yˆ (n + N | n ) = Ω × X (n ) +Ψ × Φ ( Γ × X (n ) ) (2)

where X(n) indicates the column vector of size [Nin + 1] of input
signals at time step n, plus an input equal to 1 associated with the
weight accounting for the bias:
T
 1 
X (n ) = 1, y (n ) , ∆y (n ) , RaG (n ) , (1 − z −N ) RaG (n )  (3)
 N 
Ω is the [Nin + 1] row vector of weights connecting every input

directly with the output neuron, thus Ωi = ωi indicates the weight
connecting the ith input to the output neuron. Ψ is the row vector
of size Nhn of weights connecting every hidden neuron to the out-
put neuron, thus Ψk = ψk is the weight connecting the kth hidden
neuron to the output neuron. Γ is the [Nhn × Nin] matrix of weights
connecting every input to every hidden neuron, thus Γji = γji indi-
cates the weight connecting the ith input to the jth hidden neuron.
Equation 2 can be expressed explicitly as
N in N hn
 N in 
yˆ (n + N | n ) = ∑ωi xi (n ) + ∑ψ j ϕ  ∑λ ji xi (n )  (4)
i =0 j =1  i =0 
4.3 Structure The number of hidden neurons and hidden layers was chosen with
Optimization k-fold cross-validation on the training set. Several candidate NNs
with one hidden layer and with an increasing number on hidden
neurons were evaluated. NNs with two hidden layers were also
assessed, but they did not significantly outperform NN architec-
tures with one hidden layer only. Note that the selected model is
usually a compromise between performance and structure simplicity.
Indeed the NN with the lowest possible number of neurons that
significantly outperforms simpler structures is chosen. In our case,
the best and most parsimonious architecture was a NN with one
hidden layer with five neurons. Thus, since the number of input
signals is equal to 4 and there is 1 output, there are 31 free param-
eters to tune during training.
4.4 NN Training The NN weights were optimized on the training and validation set
Procedure (see Subheading 5.1 for details) and then the architecture was
tested on an independent test set. The NN was trained with the
backpropagation Levenberg–Marquardt training algorithm,
applied in batch mode, i.e., weights and biases are only updated
after all the inputs and targets are presented to the NN, thus on the
basis of the average error obtained on the entire training set. The
minimized objective function J was the mean square error:
1 N TS 1 N TS 1 Mi
(
∑ i i i i N ∑) ( ) ∑ ( yˆ (k + N | k ) − y (k ) )
T 2
J = ˆ
y − y ˆ
y − y = i i (5)
N TS i =1 TS i =1 M i
k =1

with NTS the number of training time series and Mi the length of
the ith training time series. The training procedure was arrested
using early stopping [38]. Ordinarily, a NN learns in stages,
increasing the performance in the training set as the training ses-
sion progresses, towards a local minimum of the error surface.
However, the NN might end up overfitting the training data and

generalizing poorly. The onset of overfitting can be identified using
cross-validation: the training data are split into an effective training
set, used for computing the error and its gradient and updating the
network weights, and a validation set, used for monitoring the error
during training. The training session is stopped periodically and the
error on the validation set is computed. The validation error nor-
mally decreases during the initial phase of training; however, when
the network begins to overfit the data, the error on the validation
set begins to rise. When the error increases for a predefined number
of consecutive iterations the training is stopped and the weights at
the minimum of the validation error are returned. In our imple-
mentation, we stopped the training after 100 consecutive iterations
worsened NN performance on the validation set.
Remark: To avoid overfitting, a training procedure alternative to
cross-validation is Bayesian regularization. This technique updates
the weight and bias values according to Levenberg–Marquardt
optimization, minimizing a combination of squared errors and
squared weight values so that, at the end of training, the resulting
network has good generalization without early stopping being
required. In addition, any redundant weights in the network
should assume values close or equal to zero at the end of the train-
ing and should, potentially, be eliminated from the NN without
compromising its performance. In our implementation, this train-
ing procedure gave results comparable to those obtained with
cross-validation; however, it was considerably more time-
consuming, so we adopted the classical backpropagation with the
cross-validation algorithm. A positive feature of Bayesian regular-
ization is that it does not require any validation set, thus all the
data, other than those belonging to the test set, can be used for
training. For this reason the method might be preferable to cross-
validation when the available database is of small size.
5 Database and Assessment Criteria
5.1 Database The NN was optimized and tested on data of 20 T1D patients,
monitored for 2 or 3 consecutive days in real-life conditions. Data
were collected during the DIAdvisor™ project [39]. The glucose
concentration was measured by the Dexcom SEVEN PLUS CGM
sensor, which has a sampling time Ts of 5 min. Information on
ingested CHO was extracted by picture of meals taken by the
patients with a standard mobile phone with camera.
The database was divided into a training and validation set,
constituted by ten time series and a test set, formed by the other
ten time series. The training and validation set was further ran-
domly divided into the training set constituted by 70 % of the data,
used for minimizing the prediction error and the validation set
constituted by the remaining 30 % of data, used for stopping the
training procedure.
5.2 Assessment The quality of prediction obtained with the proposed jump NN
Metrics algorithm is quantitatively assessed by computing three metrics
commonly used in the glucose prediction literature. The consid-
ered indexes, which measure different merits of the predicted glu-
cose profile, are:
1. The Root Mean Square Error (RMSE), (mg/dl) between the
predicted time-series and the original glucose time-series mea-
sured by the CGM sensor, calculated as
2
1 M
 
RMSE =
M
∑  y (i + N | i ) − y ( i ) 
i =1  
(6)

with M being the length of the time series.

2. The average Time Gain (TG), (min)
TG = PH − delay (7)
with the delay quantified as the temporal shift that minimizes

the distance between yˆ (n + N | n ) and y(n)
 1 M −N

∑ ( yˆ (i + k | i + k − N ) − y (i ) )
2
delay = argmin  Ts (8)
k∈[0,N ]  M − N 
i =1
with Ts being the sampling period of the signal.

3. The Energy of Second Order Differences (ESOD) (i.e., the
sum of the squared second order differences) of the predicted
time series, normalized by the ESOD of the target time series
[23].
ESOD ( ŷ )
ESODnorm = (9)
ESOD ( y )
where
M
ESOD ( yˆ ) = ∑ ( yˆ (i | i − N ) − 2 yˆ (i − 1 | i − 1 − N ) + yˆ (i − 2 | i − 2 − N ) ) (10)
2

i =3
and
M
ESOD ( y ) = ∑ ( y (i ) − 2 y (i − 1) + y (i − 2 ) ) (11)
2

i =1
The RMSE is a widely used metric in the CGM literature

[22–25, 40]; however, it has some limitations: it does not
penalize spurious oscillations around the target and it is unable

to penalize differently underestimation and overestimation of
the target. TG is one of the most important indices from a prac-
tical perspective, since it quantifies the average anticipation with
which events could be, in theory, detected and thus have a clini-
cal value. The closer the TG is to the PH, the better the predic-
tion, since the patient could decide therapeutic actions ahead in
time and, likely, avoid critical events. Notably, the definition of
the delay given in Eq. 8 is consistent with those of ref. 22, 41.
ESODnorm reflects how (possibly spurious) oscillations are
amplified in the predicted time series. Thus it roughly quantifies
the risk of generating false hypo/hyper alerts. The closer to 1,
the better the predicted time series.
6 Results
Figure 3 shows prediction in one typical subject (blue line),

together with the target CGM signal (black) and meal information
provided by the patient (blue stems). The thin horizontal lines rep-
resent hypoglycemic and hyperglycemic thresholds.
The predicted signal is plotted at the time instant to which it
refers, i.e., the value plotted at a certain time is obtained N time
steps earlier, using only data available until N time steps earlier. As
we can note, the jump NN predicts the target accurately
(RMSE = 17.6 mg/dl), with a satisfactory average anticipation
(TG = 15 min) and with an acceptable amplification of measure-
ment noise (ESODnorm = 9.9). Results on the other test subjects are
in line with those of the representative subject of Fig. 3 and are
summarized in Table 1. RMSE, TG, and ESODnorm values are sat-
isfactory, suggesting that the jump NN is an appropriate architec-
ture for glucose concentration prediction and is able to learn the
Subj 1
350
CGM target
jump NN prediction
310 CHO [g]
270
230
CGM [mg/dl]
hyperglycemic threshold
180
140
100 70g CHO

70g CHO
70 hypoglycemic
40g CHO 10g CHO
40 threshold
Tue 06:00 Tue 08:00 Tue 10:00 Tue 12:00 Tue 14:00 Tue 16:00 Tue 18:00 Tue 20:00 Tue 22:00 Wed 00:00 Wed 02:00 Wed 04:00
time [Day HH:MM]
Fig. 3 CGM profile (black ) and prediction obtained with the proposed jump NN (blue line). Stems indicate CHO
ingestion, horizontal thin lines represent the hypoglycemic and the hyperglycemic threshold
Table 1
Results obtained on the ten test subjects (with PH = 30 min), average
values and standard deviation
RMSE (mg/dl) TG (min) ESODnorm (−)

Subj 1 17.6 15 9.9
Subj 2 20.3 15 10.0
Subj 3 13.1 20 8.7
Subj 4 12.0 25 7.4
Subj 5 15.6 20 8.5
Subj 6 15.9 20 12.6
Subj 7 13.7 20 9.4
Subj 8 17.8 15 8.1
Subj 9 18.2 20 12.0
Subj 10 21.3 15 9.8
Mean ± sd 16.6 ± 3.1 18.5 ± 3.4 9.6 ± 1.6
The lower the RMSE, the closer to 30 min the TG, the closer to one ESODnorm, the
better the quality of the predicted profiles
relationship between the history of CGM and of CHO ingestion

and future glucose concentration accurately. Moreover, the aver-
age TG is approximately 20 min, which is sufficient for taking
effective therapeutic decisions ahead in time.
Another interesting index is the prediction time anticipation in
correspondence of hypoglycemic and hyperglycemic thresholds. In
Fig. 3, for example, the first hypoglycemic event is not predicted,
while the second is anticipated by 5 min. Regarding hyperglyce-
mia, the first event is anticipated by 15 min and the second by
20 min. In fact, in the dataset, the number of hypoglycemic events
is limited (12 in the 10 time series used as the training set and 18 in
the 10 time series used as the test set). As a consequence, the NN
performance in forecasting these events cannot be solidly assessed.
A preliminary visual analysis, restricted to the hypoglycemic events
that the model is effectively able to predict, shows an anticipation
around 20 min in crossing the 70 mg/dl threshold. However,
some challenging events were not predicted. On the contrary,
several hyperglycemic events are present in the dataset (43 in the
10 time series used as the training set and 58 in the 10 time series
used as the test set). This allows the NN algorithm to accurately
learn them. Indeed 94 % of the hyperglycemic events are correctly
forecasted with an average anticipation of 23 min in crossing the
180 mg/dl threshold. This result is highly encouraging and sug-
gests that the NN predictor, when trained on enough representa-
tive samples, can be used to anticipate critical events.
7 Conclusions
Short-time prediction of glucose concentration can improve T1D

therapy and management; however, it is challenging since glucose
time course is affected by many exogenous disturbances (e.g.,
ingestion of CHO, injection of insulin, physical activity) and inter-
individual and intraindividual variability. The great majority of the
tens of prediction methods proposed in the literature in the last
decade are based on time series models and rarely exploit informa-
tion other than CGM. NN-based algorithms are able to learn non-
linear input–output relationships, and can easily use inputs
belonging to different domains, as information on meals, insulin
therapy and physical activity; thus they appear suitable candidate
models for predicting future glycemia for both open and closed
loop control applications.
This chapter describes a new method based on a jump NN,
with inputs directly connected both to the first hidden layer and to
the output neuron. Such a NN architecture is appropriate when
inputs and output to be learned present both linear and nonlinear
relationships. Information on timing and amount of CHO in the
ingested meal was suitably preprocessed, before entering the NN,
by exploiting a physiological model. The NN was optimized on
data of ten T1D and tested on data of ten different T1D, moni-
tored with a minimally invasive CGM sensor. Results demonstrate
that this approach is able to accurately predict glucose concentra-
tion, with an average temporal gain compatible with the genera-
tion of preventive hypoglycemic and hyperglycemic alerts.
Future work could include the investigation of other possible
inputs for the NN, such as insulin information and physical activity
related signals, whose correlation with changes in glucose dynam-
ics has recently been quantitatively investigated [42, 43]. Moreover,
the NN training procedure could be customized to minimize an
objective function that takes into account, apart from the adher-
ence to the target, also the time anticipation of prediction, possibly
penalizing differently errors in hypoglycemia, euglycemia, and
hyperglycemia.
To conclude, it is worthwhile noting that even if the presented
NN is specifically designed for glucose prediction, the methodology
could easily be adapted to the prediction of any other time series
whose dynamics are influenced both linearly and nonlinearly by
exogenous measurable signals.
References
1. De Gooijer J, Hyndman R (2006) 25 years of 2. Kruppa J, Ziegler A, Knig I (2012) Risk esti-
time series forecasting. Int J Forecasting mation and risk prediction using machine
22(3):443–473 learning methods. Hum Genet 131(7):1–16
3. Iasemidis L (2011) Seizure prediction and its 17. The American Diabetes Association (2013)
applications. Neurosurg Clin N Am 22(4): Standards of medical care in diabetes: 2013.
489–513 Diabetes Care 36(S1):S11–S66
4. Bequette BW (2010) Continuous glucose 18. Tamborlane WV, Beck RW, Bode BW (2008)
monitoring: real-time algorithms for calibra- Continuous glucose monitoring and intensive
tion, filtering, and alarms. J Diabetes Sci treatment of type 1 diabetes. N Engl J Med
Technol 4(2):404 359(10):1464–1476
5. Sparacino G, Zanon M, Facchinetti A et al 19. Battelino T, Phillip M, Bratina N et al (2011)
(2012) Italian contributions to the develop- Effect of continuous glucose monitoring on
ment of continuous glucose monitoring sen- hypoglycemia in type 1 diabetes. Diabetes Care
sors for diabetes management. Sensors 34(4):795–800
12(10):13753–13780 20. Deiss D, Bolinder J, Riveline JP et al (2006)
6. Sparacino G, Facchinetti A, Maran A et al Improved glycemic control in poorly con-
(2008) Continuous glucose monitoring time trolled patients with type 1 diabetes using real-
series and hypo/hyperglycemia prevention: time continuous glucose monitoring. Diabetes
requirements, methods, open problems. Cur Care 29(12):2730–2732
Diabetes Rev 4(3):181–192 21. Reifman J, Rajaraman S, Gribok A et al (2007)
7. Sparacino G, Facchinetti A, Cobelli C (2010) Predictive monitoring for improved manage-
“Smart” continuous glucose monitoring sen- ment of glucose levels. J Diabetes Sci Technol
sors: on-line signal processing issues. Sensors 1(4):478–486
10(7):6751–6772 22. Gani A, Gribok AV, Rajaraman J et al (2009)
8. Buckingham B, Cobry E, Clinton P et al Predicting subcutaneous glucose concentra-
(2009) Preventing hypoglycemia using prediction in humans: data-driven glucose modeling.
tive alarm algorithms and insulin pump suspen- IEEE Trans Biomed Eng 56(2):246–254
sion. Diabetes Technol Ther 11(2):93–97 23. Sparacino G, Zanderigo F, Corazza S et al
9. Buckingham B, Chase HP, Dassau E et al (2010) (2007) Glucose concentration can be predicted
Prevention of nocturnal hypoglycemia using pre- ahead in time from continuous glucose moni-
dictive alarm algorithms and insulin pump sus- toring sensor time-series. IEEE Trans Biomed
pension. Diabetes Care 33(5):1013–1017 Eng 54(5):931–937
10. Bode B, Gross K, Rikalo N et al (2004) Alarms 24. Eren-Oruklu M, Cinar A, Quinn L et al (2009)
based on real-time sensor glucose values alert Estimation of the future glucose concentra-
patients to hypo-and hyperglycemia: the guard- tions with subject specific recursive linear mod-
ian continuous monitoring system. Diabetes els. Diabetes Technol Ther 11(4):243–253
Technol Ther 6(2):105–113 25. Finan DA, Doyle FJ, Palerm CC et al (2009)
11. Garcia A, Rack-Gomer AL, Bhavaraju NC et al Experimental evaluation of a recursive model
(2012) Dexcom G4AP: an advanced continu- identification technique for type 1 diabetes.
ous glucose monitor for the artificial pancreas. J Diabetes Sci Technol 5(3):1192–1202
J Diabetes Sci Technol 7(6):1436–1445 26. Castillo-Estrada G, del Re L, Renard E (2010)
12. Cobelli C, Renard E, Kovatchev B (2011) Nonlinear gain in online prediction of blood
Artificial pancreas: past, present, future. glucose profile in type 1 diabetic patients. 49th
Diabetes 60(11):2672–2682 IEEE Conference on Decision and Control
13. Thabit H, Hovorka R (2012) Closed-loop (CDC), p 1668–1673
insulin delivery in type 1 diabetes. Endocrinol 27. Eren-Oruklu M, Cinar A, Rollins DK et al
Metab Clin North Am 41(1):105–117 (2012) Adaptive system identification for esti-
14. Zecchin C, Facchinetti A, Sparacino G et al mating future glucose concentrations and hypo-
(2014) Jump neural network for online short- glycemia alarms. Automatica 48(8):1892–1897
time prediction of blood glucose from con- 28. Turksoy K, Bayrak ES, Quinn L et al (2013)
tinuous monitoring sensors and meal Hypoglycemia early alarm systems based on
information. Comput Methods Programs multivariable models. Ind Eng Chem Res
Biomed 113(1):144–152 52:12329–12336
15. Cryer PE (2007) Hypoglycemia, functional 29. Pérez-Gandía C, Facchinetti A, Sparacino G
brain failure, and brain death. J Clin Invest et al (2010) Artificial neural network algorithm
117(4):868–870 for online glucose prediction from continuous
16. Williams G, John CP (2004) Handbook of dia- glucose monitoring. Diabetes Technol Ther
betes. Blackwell, Oxford 12(1):81–88
30. Pappada SM, Cameron BD, Rosman PM et al 37. Basheer IA, Hajmeer M (2000) Artificial neural
(2011) Neural network-based real-time predic- networks: fundamentals, computing, design,
tion of glucose in patients with insulin- and application. J Microbiol Methods 43(1):3
dependent diabetes. Diabetes Technol Ther 38. S. Amari, N. Murata, K.-R. Muller, M. Finke, H.
13(2):135–141 Yang (1996) Statistical Theory of Overtraining -
31. Daskalaki E, Prountzou A, Diem P et al (2012) Is Cross-Validation Asymptotically Effective?,
Real-time adaptive models for the personalized Advances in Neural Information Processing
prediction of glycemic profile in type 1 diabetes Systems 8, Proceedings of the 1995 Conference,
patients. Diabetes Technol Ther 14(2): edited by David S. Touretzky, Michael C. Mozer
168–174 and Michael E. Hasselmo pp 176–182
3 2. Zecchin C, Facchinetti A, Sparacino G et al 39. DIAdvisor. Personal glucose predictive diabetes
(2012) Neural network incorporating meal advisor. http://www.diadvisor.eu/. Accessed 22
information improves accuracy of Jan 2014
short-time prediction of glucose concen- 40. Gani A, Gribok AV, Lu Y et al (2010) Universal
tration. IEEE Trans Biomed Eng glucose models for predicting subcutaneous
59(6):1550–1560 glucose concentration in humans. IEEE Trans
33. McNelis PD (2005) Neural networks in Inf Technol Biomed 14(1):157–165
finance: gaining predictive edge in the market. 41. Facchinetti A, Sparacino G, Cobelli C (2010)
Elsevier Academic Press, London Modeling the error of continuous glucose
34. Dalla Man C, Camilleri M, Cobelli C (2006) A monitoring sensor data: critical aspects dis-
system model of oral glucose absorption: vali- cussed through simulation studies. J Diabetes
dation on gold standard data. IEEE Trans Sci Technol 4(1):4–14
Biomed Eng 53(12):2472–2478 42. Manohar C, Levine JA, Nandy DK et al

35. Dalla Man C, Rizza RA, Cobelli C (2007) (2012) The effect of walking on postprandial
Meal simulation model of the glucose insulin glycemic excursion in patients with type 1
system. IEEE Trans Biomed Eng 54(10): diabetes and healthy people. Diabetes Care
1740–1749 35(12):2493–2499
3 6. Facchinetti A, Sparacino G, Cobelli C
43. Zecchin C, Facchinetti A, Sparacino G et al
(2011) Online denoising method to handle (2013) Physical activity measured by physical
intraindividual variability of signal-to- activity monitoring system correlates with glu-
noise ratio in continuous glucose monitor- cose trends reconstructed from continuous
ing. IEEE Trans Biomed Eng glucose monitoring. Diabetes Technol Ther
58(9):2664–2671 15(10):836–844
Chapter 16
Preparation of Ta-O-Based Tunnel Junctions to Obtain

Artificial Synapses Based on Memristive Switching
Stefan Niehörster and Andy Thomas
Abstract
Magnetron sputtering and optical lithography are standard techniques to prepare magnetic tunnel
junctions with lateral dimensions in the micrometer range. Here we present the materials and techniques
to deposit the layer stacks, define the structures, and etch the devices. In the end, we obtain tunnel junc-
tion devices exhibiting memristive switching for potential use as artificial synapses.
Key words Memristors, Artificial synapses, Magnetron sputtering, Optical lithography, Tunnel
junctions
1 Introduction
Memristors and memristive systems have attracted a lot of interest

in recent years. The most interesting implementation of memris-
tive systems is neuromorphic devices. These aim to use biological
mechanisms operating within the brain as a prototype to construct
new computer architectures [1, 2].
A memristor is a portmanteau of memory and resistor that
Chua proposed in 1971 [3]. A straightforward way to envisage a
memristor is as an adjustable resistor, e.g., a potentiometer. If a
certain amount of flux (V × s) flows in one direction, the resistance
increases. If the current flows in the other direction through the
device, the resistance decreases. This means that the resistance of a
memristor depends on its past states, which can be used to mimic
the synaptic functionality in a brain.
In particular, it is possible to demonstrate the equivalents of
long-term potentiation (LTP), long-term depression (LTD) as
well as spike timing dependent plasticity (STDP) [4–6].
Consequently, the memristors can be used as electronic synapses
in circuits, while the integrate-and-fire functionality can easily
be achieved via conventional electronics.
261
262 Stefan Niehörster and Andy Thomas
In many cases, memristive systems consist of metal and insula-

tor thin films (~1–10 nm) [7]. Tantalum oxide based systems are
known to be stable over 1010 switching cycles [8–10], but TiO2 is
often used as well [11, 12]. We realized the memristor as a metal/
insulator/metal trilayer, in which a thin tantalum oxide layer acts
as a tunneling barrier. In the production process, the insulator
grows on top of the lower metal electrode, while the upper metal
electrode grows on top of the insulator. This asymmetry of the
manufacturing process introduces oxygen vacancies in the crystal
lattice at the bottom of the insulator. If we apply an electric voltage
across the tunnel barrier, the electric field acts on the oxygen vacan-
cies and they change their position slightly, i.e., the oxygen vacan-
cies will be shifted within the tunnel barrier. The resistance changes
considerably, because the resistance of a tunnel barrier depends
strongly on the properties (e.g., width and height) of the tunnel-
ing barrier. The oxygen vacancies remain in the tunnel contact on
their new position, which explains the memory effect. If the applied
voltage is reversed and the current flows in the opposite direction,
the oxygen vacancies also move back and, consequently, change
the resistance back to its initial value.
In the following sections, we describe the necessary techniques
and procedures to prepare tunnel junction pillars. First, we list the
required materials, i.e., the sputter targets, the substrates, and the
chemicals for the lithography and etching processes. Then, we go
into the details of the substrate preparation, the thin film deposi-
tion, the optical lithography, including ion beam etching, and
finally the production of the contact pads with insulation layer.
2 Materials
1. The substrates consist of 525 μm thick silicon wafers in <100>

orientation with a 4″ diameter. The substrates are thermally
oxidized to provide a 50 nm silicon dioxide film. The front
surface is polished, while the backside remains etched. The
wafer was doped with Boron (holes) and had a resistivity of
10–20 Ω cm. Wafers with the listed properties can be pur-
chased from, e.g., Semiconductor Wafer, Inc. We cut the wafer
into pieces of 10 × 10 mm2 for further processing.
2. Films are deposited on the wafer pieces by magnetron sputter-
ing. We utilized a “CLAB600” from “Leybold Optics, Alzenau,
Germany” with an additional oxidation chamber. The metal
films are prepared via dc- and the insulator films are deposited
via rf-sputtering.
3. We used the following sputter targets:
(a) Tantalum 3″ (4N).
(b) Palladium 4″ (3N5).
(c) Gold 2″ (4N).
Memristive Switching 263
4. The lithography was performed in a clean room equipped with

a spin-coater. We used the positive photoresist and the devel-
oper manufactured by “ALLRESIST GmbH, Strausberg,
Germany,” namely, “AR-P 5350” (resist) and “AR300-35”
(developer). The exposure was carried out by a Hg-short-arc
lamp with an emission maximum of 436 nm or by the lithog-
raphy laser system “DWL66” from “Heidelberg Instruments
Mikrotechnik GmbH, Heidelberg, Germany”, using a laser
emission maximum of 405 nm.
5. The etching of the junction pillars was completed in an argon
ion-etching chamber with a rotatable sample holder. The sys-
tem was combined with a secondary-ion-mass-spectrometer
(SIMS) to detect the depth of the etching process.
3 Methods
3.1 Substrate Substrates with a size of 10 × 10 mm2 are trimmed out of a 4″

Preparation silicon wafer. The wafer is scratched at the edge with a diamond
scraper and cut into pieces over an edge, e.g., five sheets of paper.
The substrate is affixed by a clamp to the sample holder. The sub-
strate surface has to be contacted with silver paint to the sample
holder (see Note 1). Before loading the sample into the load lock
of the sputter machine, we cleaned the sample surface with a jet of
nitrogen gas.
3.2 Thin Film 1. The thin films are deposited by magnetron sputtering. The
Deposition base pressure inside the sputter chamber is less than
2.0 × 10−7 mbar, with a set point of 2.0 × 10−7 mbar. The sputter
pressure is approximately 1.3 × 10−3 mbar and is regulated via
the Argon flow (20 sccm) and the throttle position (21 %) in
front of the turbo-pump. In our case, the tantalum film is sput-
tered from a 3″ target with a power of 65 W and the palladium
film is sputtered from a 4″ target with a power of 115 W. The
machine specific deposition rates are calibrated by X-ray reflec-
tivity measurements (0.1–0.6 nm/s).
2. We subsequently deposit a stack of 5 nm tantalum, 10 nm palla-
dium, and 2 nm tantalum on the silicon dioxide side of the wafer.
3. The sample is next moved in-situ into the oxidation chamber.
The base pressure inside this chamber is approximately
2.5 × 10−7 mbar, with a set point of 7.0 × 10−6 mbar. The
pressure during the oxidation process is approximately
2.0 × 10−3 mbar and depends on the oxygen flow (13 sccm) and
the throttle position (80 %) in front of the turbo-pump. The
microwave power to ionize the oxygen is set to 275 W and the
bias voltage to accelerate the oxygen ions towards the sample
is −80 V. The sample is oxidized for 150 s to obtain the tanta-
lum oxide tunnel barrier.
4. The sample (in-situ) is then moved back into the sputter cham-
ber and layers of 10 nm Ta, 5 nm Pd, 5 nm Ta, and 30 nm Au
are deposited. Gold needs a higher sputter pressure of approxi-
mately 4.5 × 10−3 mbar, which is realized by a smaller opening
of the throttle (8 %). The sputter power for gold is 29 W,
because of the smaller target size of 2″.
3.3 Optical This is carried out as follows:

Lithography
1. Place the sample in the middle of the spin-coater and tape it
with double-sided sticky tape. Fill a one-way pipette with pho-
toresist and clean the sample with a jet of nitrogen gas. Drop
the photoresist on the sample until the surface is fully covered.
Rotate the sample at 4,000 rpm (or 6,000 rpm for the laser
lithography) and subsequently bake it at 80 °C for approxi-
mately 30 min.
2. Place a mask with an adequate layout on the sample and irradi-
ate both with the UV-light (Fig. 1). For positive resist cover
the junction pillars. Alternatively, use the DWL-66 laser lithog-
raphy to create the desired layout. Our layout consists of
squares with sizes of 7.5 × 7.5 μm2, 12.5 × 12.5 μm2, and
22.5 × 22.5 μm2.
3. Mix the photo developer with pure water at a ratio of 2 to 1.
Stir the sample in the mixture for about 8 s until the vapor
dissolves (see Note 2), then stop the developer process and
clean it with pure water. Afterwards, dry the sample with
nitrogen gas.
Fig. 1 (a) The complete mask layout. (b) The upper right edge of the layout design. The black parts cover the
sample (positive resist). The triangles in the frame are markers for easier positioning, in particular, if a second
lithography step is necessary. The mask has to be placed upside down on the sample, so the characters are
reversed to get a correct layout on the sample
3.4 Ion Beam 1. Clean dust specks from the sample with nitrogen gas and lock
Etching it in the etching chamber.
2. The base pressure inside the etching chamber should be
approximately 3.0–4.0 × 10−8 mbar and the etching (sputter)
pressure approximately 3.5 × 10−5 mbar. The pressure is regu-
lated by the Argon flow (approximately 2.0 sccm). The cath-
ode current should be in the range 300–400 μA. The sample
holder should rotate to achieve an equal etching rate over the
entire sample. We have to etch through the tantalum oxide
barrier into the palladium film (Fig. 2) (see Note 3).
3. After the etching procedure, it is possible to remove the
remaining squares of photoresist through the use of an ultra-
sonic bath in acetone for approximately 5 min (see Note 4) and
in ethanol for approximately 1 min. Only remove the photore-
sist, if no contact pads are needed and skip Subheading 3.5 as
well as 3.6.
Fig. 2 The main steps from UV-exposition to the finished sample

3.5 Insulator 1. Before removing the photoresist, we mark one sample edge
with a felt pen. Fill the spaces between the junction squares
with an isolator such as silicon nitride or tantalum oxide via
rf-sputtering. The insulator should have at least the same thick-
ness as the sum of the etched films (here ≥60 nm).
2. Now, remove the photoresist as well as the felt pen mark with
acetone in an ultrasonic bath. In our case, the sample has to
stay 5–10 min in the ultrasonic bath in a beaker filled with
acetone because of the processed photoresist (see Note 4).
3. The felt pen mark covered the surface and allows an easy
removal of the insulator on top of it, therefore, providing
access to the lower electrode.
4. This procedure makes it easier to contact the elements without
a short circuit and protects the sample from scratches and
other damage or adverse environmental conditions.
3.6 Contact Pads It is possible to deposit contact pads on the elements, if the sizes of
the junction pillars become very small. While it is easy to contact a
junction pillar with lateral dimensions larger than 100 μm, it is very
challenging to do this if the sizes approach 10 μm. In our case, we
added contact pads with a size of 50 × 100 μm2 (see Notes 5 and 6).
1. At first, it may be necessary to spin coat the sample with pho-
toresist as described in Subheading 3.3, step 1.
2. The contact pads are prepared using a lift-off process. Therefore
we need a mask with a layout matching the previously used
junction mask, i.e., the whole sample is covered and the
50 × 100 μm2 holes are on top of the junction pillars. Place the
mask on the sample and irradiate it with the UV-light.
3. Sputter a 5 nm film of tantalum and a 60 nm film of gold, as is
described in Subheading 3.2, steps 1 and 4.
4. Place the sample in a beaker filled with acetone for the lift-off
process and place it for 5–10 min in the ultrasonic bath (similar
to Subheading 3.5)
4 Notes
1. The sample is connected with silver paint to the sample holder.

This is very important for the oxidation process, because it
prevents charging of the sample.
2. According to the age of the resist and the developer, the size of
the sample and of the junction pillars, the developing time may
vary about some seconds.
3. The exact etching depth is not important. However, we have
to meet two requirements. First, we must etch through the
barrier to prevent a short-circuit of the junctions via the top

metal layers. Secondly, we must stop before the lower elec-
trode thickness falls below approximately 10 nm, otherwise,
the resistance of the lower electrode becomes higher than the
resistance of the tunneling barrier which may lead to artifacts
in the measurements.
4. The time in the ultrasonic bath is determined by the size of the
junction pillars. Smaller junctions require longer removal, e.g.,
30 min for sizes below lateral sizes of 1 μm. The removal of the
resist cap should be checked with an optical microscope or a
scanning electron microscope for junction sizes below 0.5 μm.
5. If contact pads are used, it is essential to increase the distance
of the junction pillars in the design of the first mask. Otherwise,
the larger contact pads will overlap and short-circuit neighbor-
ing junctions.
6. The contact pads should be placed off-center, with one edge close
to the junction pillar. This prevents the junctions from damage, if
we want to contact them by for example wire-bonding.
References
1. Thomas A (2013) Memristor-based neural net- tunnel junctions. Appl Phys Lett
works. J Phys D Appl Phys 46(9):093001. 95(11):112508. doi:10.1063/1.3224193
doi:10.1088/0022-3727/46/9/093001 8. Yang JJ, Zhang MX, Strachan JP et al (2010)
2. Turel O, Lee J, Ma X et al (2004) Neuromorphic High switching endurance in TaOx memristive
architectures for nanoelectronic circuits. Int J devices. Appl Phys Lett 97(23):232102.
Circ Theor Appl 32(5):277–302. doi:10.1002/ doi:10.1063/1.3524521
cta.282 9. Strachan JP, Torrezan AC, Medeiros-Ribeiro
3. Chua L (1971) Memristor: the missing circuit G et al (2011) Measuring the switching
element. IEEE Trans Circ Theor 18:507–519 dynamics and energy efficiency of tantalum
4. Krzysteczko P, Münchenberger J, Schäfers M oxide memristors. Nanotechnology
et al (2012) The memristive magnetic tunnel 22(50):505402. doi:10.1088/0957-4484/
junction as a nanoscopic synapse-neuron sys- 22/50/505402
tem. Adv Mater 24:762–766 10. Torrezan AC, Strachan JP, Medeiros-Ribeiro G
5. Afifi A, Ayatollahi A, Raissi F (2009) STDP et al (2011) Sub-nanosecond switching of a
implementation using memristive nanodevice in tantalum oxide memristor. Nanotechnology
CMOS-Nano neuromorphic networks. IEICE 22(48):485203. doi:10.1088/0957-4484/
Electron Exp 6(3):148–153. doi:10.1587/ 22/48/485203
elex.6.148 11. Yakopcic C, Sarangan A, Gao J et al. (2011)
6. Indiveri G, Chicca E, Douglas R (2006) A TiO2 memristor devices. IEEE National
VLSI array of low-power spiking neurons and Aerospace & Electronics Conference, July
bistable synapses with spike-timing dependent 2011, p 101–104
plasticity. IEEE Trans Neural Netw 17(1):211– 12. Pickett MD, Strukov DB, Borghetti JL et al
221. doi:10.1109/TNN.2005.860850 (2009) Switching dynamics in titanium dioxide
7. Krzysteczko P, Reiss G, Thomas A (2009) memristive devices. J Appl Phys 106(7):074508.
Memristive switching of MgO based magnetic doi:10.1063/1.3236506
Chapter 17
Architecture and Biological Applications of Artificial

Neural Networks: A Tuberculosis Perspective
Jerry A. Darsey, William O. Griffin, Sravanthi Joginipelli,
and Venkata Kiran Melapu
Abstract
Advancement of science and technology has prompted researchers to develop new intelligent systems that
can solve a variety of problems such as pattern recognition, prediction, and optimization. The ability of the
human brain to learn in a fashion that tolerates noise and error has attracted many researchers and pro-
vided the starting point for the development of artificial neural networks: the intelligent systems. Intelligent
systems can acclimatize to the environment or data and can maximize the chances of success or improve
the efficiency of a search. Due to massive parallelism with large numbers of interconnected processers and
their ability to learn from the data, neural networks can solve a variety of challenging computational prob-
lems. Neural networks have the ability to derive meaning from complicated and imprecise data; they are
used in detecting patterns, and trends that are too complex for humans, or other computer systems.
Solutions to the toughest problems will not be found through one narrow specialization; therefore we
need to combine interdisciplinary approaches to discover the solutions to a variety of problems. Many
researchers in different disciplines such as medicine, bioinformatics, molecular biology, and pharmacology
have successfully applied artificial neural networks. This chapter helps the reader in understanding the
basics of artificial neural networks, their applications, and methodology; it also outlines the network learn-
ing process and architecture. We present a brief outline of the application of neural networks to medical
diagnosis, drug discovery, gene identification, and protein structure prediction. We conclude with a sum-
mary of the results from our study on tuberculosis data using neural networks, in diagnosing active tuber-
culosis, and predicting chronic vs. infiltrative forms of tuberculosis.
Key words Artificial neural networks, Intelligent systems, Bioinformatics, Drug discovery, Pattern
recognition, Tuberculosis, Protein structure prediction, Gene identification
1 Introduction
Artificial neural networks (ANNs) are biologically inspired com-

puter programs designed to simulate the way in which the human
brain processes information [1]. ANNs are inspired by the way in
which a biological nervous system, such as the brain, processes the
information. The information paradigm in ANNs gathers knowl-
edge from learning experience (through inspection of training
269
270 Jerry A. Darsey et al.
data) and makes intelligent predictions on the specified data set.

An ANN is formed through the combination of a few to as many
as hundreds of single units, artificial processing elements such as
nodes which are connected with coefficients, or weights of process-
ing elements and is organized into layers. A simple ANN has input,
hidden, and output layers, while more complex networks can have
more than one hidden layer with one input and one output layer.
ANNs have the ability to derive meaning from complicated and
imprecise data and can extract patterns and detect trends that are
too complex for the human brain or other computer techniques to
recognize. Neural networks are different from conventional com-
puter techniques, as they encompass adaptive learning and real-
time operation, whereas conventional computer techniques use
algorithmic and programming approaches.
1.1 Human Brain The average human brain has about 100 billion neurons or nerve
cells, and a slightly greater number of neuroglial cells, which support
and protect neurons. Neurons are considered to be the information
transmission units of the brain, and each neuron is connected to up
to 10,000 other neurons [2]. Neurons are the basic units of the
nervous system, and have a cell body or soma, together with den-
drites (inputs) and axon (outputs), as shown in Fig. 1. Each neuron
is connected to other neurons through a synapse; typically a synapse
is filled with a neurotransmitter chemical and connects the axon of
one neuron and the dendrite of a second neuron. Signals are trans-
mitted from one neuron to another through synaptic space; these
signals can be either excitatory or inhibitory. Neurons are organized
and extensively connected to form a complex network in the human
brain, and act as messengers in sending and receiving impulses (sig-
nals). The complex network makes the human brain intelligent, with
learning, memory, recognition, and prediction capabilities.
Fig. 1 Basic structure of neuron with cell body or Soma, dendrites as input nodes,
and axon as output nodes. Branches from axon are connected to dendrites of
other neuron in complex network of human brain
Artificial Neural Networks Applied to TB 271
1.2 Architecture Although the basic unit in the structure of an ANN is computa-
of Artificial Neural tionally similar to the structure of a neuron in the human brain, it
Networks is much simpler biologically than a neuron in the brain. ANNs
incorporate two fundamental elements of biological neural net-
works; neurons are represented as nodes and synapses are repre-
sented as weighted interconnections [3]. Artificial neurons are
organized into layers: the input layer, hidden layer(s), and an out-
put layer. Neural networks with one hidden layer are sufficient to
make decisions and predictions on simple non-linear approxima-
tions, but networks with two or more hidden layers are used in
more complex applications. A Neural network with one hidden
layer is known as a simple perceptron, and with more than one hid-
den layer is called a multilayer perceptron [4]. As shown in Fig. 2,
the input layer is made up of nodes, which contain an activated
function. This layer communicates with one or more hidden layers,
where the actual processing is done, through weighted connec-
tions. The final hidden layer then links to the output layer, which
generates the network outcome or predictions.
x1, x2, and x3 …. are inputs and w1, w2, and w3 … are weighted
connections expressing the importance of inputs to the outputs
(Fig. 1). The network’s output can in principle be any value, but in
many applications is limited to the binary set {0, 1} This output is
W1
Hidden Layer
Input layer
Output Layer
X1
X2
X3
W3 W2
Fig. 2 Architecture of artificial neural network or simple perceptron with input,

hidden, and output layers. x1, x2, and x3 of input layer corresponding to activation
function and w1, w2, and w3 interconnections corresponding to weights
Fig. 3 Architecture of neural networks based on the connection patterns
determined by whether the weighted sum Sw x

j
j j is less than or
greater than a threshold value [5]:
ì0 if
ï
åw x
j
j j £ threshold
output = í
ï1 if åw x j j > threshold
î j

The working mechanism of ANNs
Artificial neural networks are considered as learning algorithms
with nodes as activation functions and weights as edges or connec-
tions between the nodes. Based on the connection pattern or
architecture of the network, ANNs can be either feed-forward net-
works or recurrent (feedback) networks (Fig. 3).
1.2.1 Feed-Forward Feed-forward networks can be either single- or multilayer percep-

Network trons with unidirectional connections, and are generally static and
memory-less systems [5, 6]. In feed-forward networks the flow of
information is in one direction only, with no loops, thus leading to
only one set of output values. As the response to input is indepen-
dent of the previous state, they are considered as memory-less
systems [6].
1.2.2 Feedback Network Feedback networks are dynamic, and the output is dependent on
or Recurrent Network the previous layer.
When a new input pattern is presented, the neuron outputs are
computed. Because of the presence of feedback paths, the inputs to
each neuron are then modified, which leads the network to enter a
new state [6].
1.3 Learning The learning process in ANNs can be viewed as the process of
Algorithm of Neural updating network architecture and adjusting weighted patterns
Networks to efficiently achieve a specific task. A network learns from avail-
able training patterns, and the performance is improved over
time by iteratively adjusting weight patterns in the network. The
major advantage of neural networks when compared to tradi-
tional expert systems is that they learn from a given set of repre-
sentative examples of training data rather than by following a set
of rules previously defined by human experts, as in traditional
computer systems [6].
The learning process or learning architecture of ANNs can be
designed from known, available information about the training
data set. A learning paradigm can be defined by having a model
from a known environment, and learning rules can be defined by
figuring out the update process for connection weights [6].
Identifying a procedure to adjust weights by use of a learning rule
and following a learning paradigm will define the learning process
or learning algorithm of the network.
The major learning paradigms of ANNs are
1. Supervised.
2. Unsupervised.
3. Hybrid.
In a supervised learning paradigm, a correct answer is provided
for each input pattern and the weights are adjusted based on the
degree to which network output agrees with that answer. In an
unsupervised paradigm, as the answers are not provided, the sys-
tem itself recognizes a correlation and organizes the patterns into
categories accordingly. The hybrid paradigm combines both super-
vised and unsupervised learning methods. In hybrid learning
methods, some weights are provided with correct answers, while
some are auto corrected.
The four basic learning rules for a learning paradigm of
ANNs are
1. Error-correction rules.
2. Boltzmann learning rule.
3. Hebbian rule.
4. Competitive learning rules.
1.3.1 Error- All these learning rules can be applied to supervised, unsupervised,
Correction Rule and hybrid paradigms. In the error-correction rule learning pro-
cess, the error generated in each step is used to adjust the weights
to efficiently achieve a task [6]. In the learning process the actual
output y is different from the desired output d, so an expression
related to the error (d−y) is used to adjust the weights to gradually
reduce the error.
1.4 Boltzmann Boltzmann learning is used in symmetric recurrent networks, i.e.,

Learning Rule those in which wij = wji, and consists of binary function units such
as +1 for on and −1 for off. Outputs in this rule are produced
according to Boltzmann statistical mechanics. Boltzmann learning
adjusts weights until the visible units satisfy a desired probabilistic
distribution [6].
1.4.1 Hebbian Rule Hebbian rule is based on neurobiological experiments and is the
oldest learning rule [6]. An important property of this rule is that
learning is done locally; that is, change in weight depends only on
the activities of two layers or neurons connected by it. Orientation
selectivity occurs from a Hebbian training network.
1.4.2 Competitive The basis of competitive learning is the “winner-take-all” process

Learning Rule observed in biological neural networks. All input units are con-
nected together and all units of output are also connected via
inhibitory weights; however, the latter are fed back with excitatory
weights. As a result of this learning process the pattern in the win-
ner unit (weight) becomes closer to the input pattern.
2 Biological Applications of ANNs
ANNs are extremely powerful systems, with massive parallelism.

They can learn and generalize from training data; the amount of
programming required to efficiently achieve the task is usually
modest. They are noise and fault tolerant, so they can cope with
situations in which more conventional systems have difficulty. The
inherent ability of ANNs to learn and recognize highly nonlinear
and complex relationships in data makes them ideally suited to a
wide range of applications. ANNs are useful in inferring function
or predicting function from a set of observations where the com-
plexity of data suggests that the use of conventional algorithms is
unlikely to be successful. ANNs are used in regression analysis,
function prediction, classification of patterns, character recogni-
tion, and image compression. Neural network applications are
used in many disciplines such as medicine, bioinformatics, molecu-
lar biology, and pharmacology. ANNs are currently used in medi-
cine for medical diagnosis, biochemical analysis, image analysis,
and drug discovery [7].
Medical diagnosis: Due to the ability of neural networks to learn
and recognize patterns from data, ANNs are currently used in
medical diagnosis of cancer, to predict the outcome of chemother-
apy in various stages of cancer research [7].
Biochemical analysis: ANNs have been used to analyze blood, urine
samples, and body fluids. They can track the level of glucose in
diabetes, and track the level of proteins in body fluids to detect
abnormalities in disease conditions. They can identify drug resis-

tance in pathological conditions such as tuberculosis [7].
Image analysis: Medical image data-generated ultrasound, MRI,
and X-rays can be analyzed with ANNs to detect disease condi-
tions. Electrocardiographic (ECG) data can be analyzed to identify
possible cardiovascular disease [7].
Drug discovery: Neural networks are used to identify potential drug
candidates by virtual screening of large number of molecules of poten-
tial value in the treatment of AIDS, cancer, and tuberculosis [7].
ANNs are used in molecular biology and bioinformatics, for
recognition of transcription and translational signals, protein struc-
ture prediction, gene identification, and sequence classification [8].
Recognition of transcription and translational signals: Recognition
of transcription and translation promoter regions and termination
of transcription and translational signals are useful in understand-
ing defects in protein synthesis and defects in protein folding in
inherited diseases such as Parkinson’s disease and multiple sclero-
sis. Kalate, Tambe, and Kulkarni used ANNs to predict mycobacte-
rial promoter sequences, and were able to achieve 97 % success in
prediction using a multilayered neural network, trained with error
back-propagation [8]. Scheila de Avila and Günther have used
neural networks to predict prokaryotic promoter sequences and
were able to achieve good results with high accuracy [9]. Zhang
et al. successfully predicted promoter regions in Escherichia coli
using a feed-forward neural network [10].
Protein structure prediction: Protein structure predication involves
identification of secondary structure and tertiary structure of a
protein, and identification of three classes of secondary structures:
α-helix, β-sheets, and random coils. Researchers have successfully
applied neural networks to the prediction of protein structure.
Holley and Karplus have used neural networks to predict protein
structure, and achieved a maximum overall predictive accuracy of
63 % for helix, sheet, and coil. When filtering was used to include
only the strongest 31 % of predictions, the predictive accuracy rose
to 79 % [11].
Gene identification: Gene identification is simple in prokaryotes
because the coding region in prokaryotes is a single strand of con-
tinuous reading frame, whereas in eukaryotes it consists of introns
and exons. Therefore, an important task in identification of the
gene-coding region in eukaryotes involves differentiation of introns,
exons, and splice sites. The ANN approach has been applied to the
recognition of coding region and gene identification; suitable gene
identification tools and software have been developed by research-
ers. GRAIL and NetGene are gene identification software tools
developed using ANNs. GRAIL is a multi-agent neural network
system for gene identification developed by Ying Xu et al. [12].
NetGene server is a service producing neural network predictions

of splice sites in human, C. elegans, and A. thaliana DNA [13, 14].
NetGene server is available online at URL http://genome.cbs.dtu.
dk/services/NetGene2/
Sequence classification: A three-layered, feed-forward, back-
propagation neural network was used by Wu et al. for full-scale
protein sequence classification; using sequence encoding with sin-
gular value decomposition, they achieved close to 90 % sensitivity
[15]. Blekas et al. have used ANNs for motif-based sequence clas-
sification; their experimental results on real datasets indicated that
the ANN method was highly efficient and superior to other well-
known protein classification methods [16].
We have successfully applied a neural network approach in our
research on tuberculosis (TB), and Parkinson’s disease. ANNs were
used for identification of TB+ and TB−, and for the prediction of
chronic and infiltrative TB from a given dataset. Neural networks
combined with molecular modeling studies on Geldanamycin and
similar compounds have predicted potential drug candidates for
treating Parkinson’s disease. Results generated from ANN applica-
tions on TB are shown in the results section of this chapter.
3 Methodology
It is believed that knowledge in the brain is gained by constant

adaptation of synapses to different input signals, yielding better
output signals; the results are constantly fed back as new inputs. In
a similar fashion, ANNs try to mimic the adaptation of synapses by
iteratively adjusting the weights associated with each node accord-
ing to the differences between desired output and actual or
obtained output.
As mentioned in Subheading 1, the ANN can have three or
more layers and the results generated in each layer are back propa-
gated to minimize the error. Figure 4 shows a sample of how the
formulae coalesce into a neuron during training on a set of incom-
ing data, passing through a hidden layer of nodes, and predicting
an output value [17]. In all but the input layer the nodes compute
the sum of the product of the input values passed on by the previ-
ous layer and the weight of the connection. After processing by a
squashing or activation function, the value is sent to nodes in the
next layer. The simplest configuration has one hidden layer with
multiple nodes, but so-called deep networks have multiple hidden
layers, to find complex interrelationships of input values [18]. An
ANN with one hidden layer can be used to approximate any func-
tion, as shown in the Cybenko theorem [19].
Selection of a training set is an important step in training
ANNs. The dataset is divided into three portions: a control group,
NODE
{
m-1
V1
m-1
V2 OUTj
INPUT VECTOR V •
S F
•
•
m-1
Vn
m
biasj
m-1 NETj
m m m-1 m-1
NETj = S Wij Vi +qi
m = LAYER NUMBER
W = WEIGHT
OUTj =F(NETjm) F = TRANSFER FUNCTION
m 1
F(NETj )= m
1+ exp(-NETj )
Fig. 4 The diagram of the node and transfer function of an artificial neural
network
a training set, and a test set. Each of these portions contains

approximately one-third of the data, with the training set often the
smallest and the test set the largest. Control data is used as a sup-
plier of feed information for correcting the ANN architecture for
the learning approach. The training in an artificial neural network
proceeds by passing values from the training set through the net-
work many times, and inspecting the network outcome. The con-
nection weights are adjusted as training proceeds as training seeks
to minimize the error in the output. There are also parameters that
govern an overall rate for learning and a momentum for the train-
ing which makes adjustments to how the weights are altered. A bias
(or threshold) value regulates the prediction around a value, which
is shown in Fig. 4.
The data which the neural network generalizes is typically pre-
normalized to values between zero and one. It is also critical for
the data to avoid the extremes of a sigmoidal curve; typical ANN
practice uses values near zero and one, i.e., 0.05 and 0.95, to
replace outputs of less than 0.05 and greater than 0.95, respec-
tively. This constraint limits the data which can be predicted or
analyzed. One consequence is that data values may need to be
altered, for example by taking the logarithm, if values within the
dataset lie outside these limits. A Boolean data type (true or false)
can also be predicted using these numeric values as well by round-
ing the network output to integer values.
The quality of an ANN prediction can be measured in its adapt-
ability, its ability to recognize patterns, and low generalization error.
This greatly reduces the computational time for large datasets.
All prediction networks, which include artificial neural networks,

are susceptible to factors which can limit their success, including
overtraining, there being too few examples in the data set to train
effectively, unreliable or noisy data, or missing factors needed for
the prediction. Overextended or convoluted predictions are avoided
by limiting the number of hidden nodes and the number of cycles
for training the neural networks. Furthermore, the order in which
data are presented to the network is randomized each training cycle
to minimize selection bias in the test set and training set.
The learning algorithm computes good predictions by adjust-
ing weights in a manner that keeps the network predictions simple
and straightforward. Cross-validation is needed to check for over-
training and generalization errors [20]. The cross-validation is
done with the leave n out technique, where n test cases are removed
from the dataset in order to see if the neural network has learned
from the data set. The number n can vary from 1 to as much as
about 10 % of the total data available for training. Passing the
value(s) withheld to the neural network, its prediction also improves
over several learning cycles.
A partial least squares linear regression model applied to the
data may help to validate the correlation found in the ANN model
[21, 22]. Other tests available to researchers to confirm ANN are
the runs test [22], which examines the number of series of con-
secutive values in the data for nonlinearity; Mallows augmented
partial residuals plot [23, 24], a universal diagnostic tool which
finds nonlinearities; Durbin-Watson test [25], which looks at the
null hypothesis that there is no correlation; principal component
analysis [26], a transformation procedure of the data into linear
correlated parameters and scored components; and partial least
squares analysis [26], a linear regression method which projects
predicted and observable variables into a new space.
The program we used for prediction, developed at NASA’s
Johnson Space center, is called NETS [27]. The data used in this
study was acquired from patients at the Kyrgyzstan National
Center for Tuberculosis Research [28]. The presence or absence
of genetic factors and types of TB is displayed in the data set as
well as a few physiological factors: blood type, Rh factor, sex, age
range, and northern or southern dwelling Kyrgyz. There were
149 patients in the control group and 146 in the test group.
Although some of the occurrences of types of TB were rare
within the test group, all patients were assessed in the
ANN. Normally outliers in the prediction would be removed if
warranted and the dataset examined again without the outliers.
Since the distribution is bimodal, we were double blinded from
examining results or patients, and the data set is from a small
portion of the TB population from one country; the data were
not corrected for outliers.
4 Results
As an example of biomedical applications of ANNs, we present

results of a study that we performed on prediction of active TB from
genetic factors in TB patient data. The network was predictable for
many of the types of TB where sufficient training examples were
available in the data set. At a plateau in the data, the network achieved
stable training over several cycles. We used the results early on in
these areas to extract weight values. In an analysis of the resulting
weights, weight values lower than the mean were removed and the
factors with significant influence on the outcome are shown in the
inset of Fig. 5. Promotion or inhibition of the TB is given as a
weighted value of the factors in the positive or negative direction.
The network predicted TB correctly in 72 % of the patients
after ~145,000 cycles as seen in Fig. 5. The results revealed certain
genes that inhibited or promoted TB likelihood. The top half of
the weight values over the mean are shown in the inset as genes
which promoted or inhibited the prediction.
A second network which predicted both the presence and
absence of TB plateau for both cases after ~85,000 cycles is shown
in Fig. 6. The resulting weight values were extracted and the top
half of the weight values over the mean are shown in the inset of the
graph. The weight values mirror each other exactly as one would
expect when predicting both causative and inhibitory factors.
The third network was trained to predict chronic and infiltra-
tive TB cases in the data set as seen in Fig. 7. The chronic type of
Accurancy of TB+ ANN

80
70
60
Selection Factors Identified by Configuration 1
Percentage Correct
50 60
40
40
20
0
30
G3 G5 G10 G13 G14 G15 G42 G47 G50 G51 G54 G57 G58 G72 G80 G81 G82 G83
-20
20 -40
-60
10 -80
0
0 200 400 600 800 1,000 1,200 1,400 1,600 1,800 2,000 2,200 2,400
Cycle/100
Fig. 5 The neural network prediction of TB cases in the dataset. Inset figure shows selection factors for the
prediction which promoted or inhibited prediction by their relative weight strength
Accuracy of Tb+/Tb- ANN

90
80
70
60
Percentage Correct
50 Selection Factors Identified by Configuration

100
40 80
Series s2
60
40
30 20 Series s1
TB-
0
TB+
−20 G5 G9 G10 G13 G14 G30 G45 G46 G47 G50 G51 G57 G68 G69 G72 G81 G82
20 −40
−60
−80
10 −100
0
0 20 40 60 80 100 120
Cycle/1000
Fig. 6 Prediction of TB-infected and control group (non-TB) patients with ANN. The inset shows the selection
factors and the weight strength of those factors
Accuracy of Chronic vs. Infiltrative ANN

Prediction
90
80
70
60
Percentage Correct
50 Selection factors identified for Infiltrative and Choronic TB

60
40 40 chronic
20
0 Infiltrative Infiltrative
30 −20 chronic
−40
20
AB
G01
G03
G07
G08
G10
G13
G14
G26
G38
G40
G42
G44
G45
G46
G50
G51
G54
G56
G57
G58
G72
G86
male
−60
10
0
0 20 40 60 80 100 120
Cycle/1000
Fig. 7 The prediction of chronic and infiltrative TB from the data set. Inset shows factors which promoted or
inhibited prediction of each type of TB
TB contained data for patients labeled: chronic, FKT, and empyema.

The infiltrative type of TB contained the categories: infiltrative,
positive test for mycobacterium TB, and drug-resistant TB. After
~18,000 cycles, the plateau region gave 77 % predictability. The
inset figure shows the weight values which promote or inhibit both
kinds of TB.
5 Discussion and Conclusions
The most important aspect of a TB infection control program is to

identify patients with active TB; they can be isolated, as active TB is
contagious, and can be efficiently treated. Early diagnosis is the key
to effectively treating and preventing infection. The ANN method
can be applied to diagnose active TB. Drug resistance is the most
alarming and growing threat to the effective treatment of
TB. Identification of drug resistance takes from several days to a few
weeks, delaying the ability to treat a patient in the early stages.
Predictive models such as neural networks can be used to identify
drug resistance in mycobacterium; this could reduce the time required
to diagnose drug resistance by traditional culture methods.
Radiological and clinical information from suspected patients
can be used to diagnose TB, and that will be superior to a physician’s
opinion. A multilayered network approach could be used to diag-
nose TB, and to detect drug resistance. Computational methods and
predictive models such as neural networks or hidden Markov models
could reduce the time and cost associated with diagnosis and detec-
tion of lethal and infectious diseases [29]. The outcome of treatment
by chemotherapy and radiotherapy for cancer and cardiovascular dis-
eases can be predicted with neural networks.
The factors which promote and inhibit TB susceptibility are
tools for researchers interested in tracing the genetic pathways of
TB infection. Doctors with this information could more accurately
identify patients who may be infected but as the waxing and wan-
ing of the symptoms progress, they are unaware of their infection.
The genetic factors which inhibit TB could be used to study resis-
tant populations and help find ways to find effective vaccines for
TB. Identification is a key step in research, because as the myco-
bacterium tuberculosis develops more drug-resistant forms, the
search for vaccines becomes more critical. The triggers for TB are
unknown, and more research into those triggers would help iden-
tify more TB risk factors. Because genetic and physical factors are
unalterable by the patients and environment, they can be attrib-
uted to the infection directly and experimented in a more con-
trolled manner than in a study of external factors.
The first two predictions are comparable in their scope by
predicting in the control versus infected patients. Both give the
same percent predictability at nearly 70 % of the withheld group.
The data in the second prediction (TB+/TB−) has an inset of

selection factors which mirror the results obtained by both tests.
The results show that the associations the ANN link to the pre-
dicted parameter are not random and corroborate the promotion
and inhibition factors consistently.
The third ANN prediction found parameters which affect the
chronic and infiltrative types of TB. In addition to genetic links, it
found links in the AB blood type, and male sex for chronic
TB. There are certain genetic parameters which oppositely reflect
both cases, either inhibiting or promoting the cases of chronic and
infiltrative TB (genes 1, 8, 13, 42, and 56 in inset of Fig. 7). These
oppositely reflecting genes may hold keys which identify the type
of TB a patient would be likely to contract.
Improvements can be made in how the data is used. The TB
and non-TB patients group each had almost 150 patients per
group. This sample size would be improved by removing outliers
to the prediction, and with 150 cases in each group, this is a large
enough study to prune outliers. The number of input factors could
also be trimmed since there are some factors which have more
influence than others.
References
1. Agatonovic-Kustrin S, Beresford R (2000) 8. Kalate RN, Tambe SS, Kulkarni BD (2003)
Basic concepts of artificial neural network Artificial neural networks for prediction of
(ANN) modeling and its application in phar- mycobacterial promoter sequences. Comput
maceutical research. J Pharm Biomed Anal Biol Chem 27(6):555–564, ISSN 1476-9271
22(5):717–727, ISSN 0731-7085. http:// 9. de Avila S, Silva GJLG (2011) Rules extraction
www.sciencedirect.com/science/article/pii/ from neural networks applied to the prediction
S0731708599002721 and recognition of prokaryotic promoters.
2. MacDonald M (2008) Your brain: the missing Genet Mol Biol 34(2):353–360
manual. Pogue, Sebastopol, CA 10. Zhang F, Kuo MD, Brunkhors A (2006) E. coli
3. Gonzalez AJ, Dankel DD (1993) The engi- promoter prediction using feed-forward neural
neering of knowledge-based systems. Prentice- networks. Proceedings of the 28th IEEE, EMBS
Hall, Englewood Cliffs, NJ. ISBN 0-13- Annual International Conference, New york
334293 City, USA 2025–2027
4. Zahedi Z (1993) Intelligent systems for busi- 11. Holley HHL (1989) Protein secondary struc-
ness: expert systems with neural networks. ture prediction with a neural network. Proc
Wadsworth Inc., Belmont, CA. ISBN 0-534- Natl Acad Sci U S A 86(1):152–156
18888-5 12. Xu Y, Mural RJ, Einstein et al (1996) GRAIL:
5. Nielsen MA (2014) Neural networks and deep a multi-agent neural network system for gene
learning, Chapter 3, Improving the way neural identification. Proc IEEE 81(10):1544–1552
networks learn. Determination Press. http:// 13. Hebsgaard PGK, Tolstrup N, Engelbrecht J
neuralnetworksanddeeplearning.com/chap1. et al (1996) Splice site prediction in Arabidopsis
html thaliana DNA by combining local and global
6. Jain AK, Mao JC, Mohiuddin KM (1996) sequence information. Nucleic Acids Res
Artificial neural networks: a tutorial. Computer 24(17):3439–3452
29(3):31 14. Brunak S, Engelbrecht J, Knudsen (1991)
7. Dybowski R (2000) Neural computation in Prediction of human mRNA donor and accep-
medicine: perspective and prospects. Artificial tor sites from the DNA sequence. J Mol Biol
neural networks in medicine and biology. 220:49–65
Proceedings of the ANNMAB-1 conference, 15. Wu C, Shivakumar S, McLarty J (1995)

Sweden, ISBN 978-1-85233-289-1 Neural networks for full-scale protein
sequence classification: sequence encoding pharmaceutical samples. Anal Bioanal Chem

with singular value decomposition. Machine 374:460–465
Learning 21(1–2):177–193 23. Centner V, de Noord OE, Massart DL (1998)
16. Blekas K, Likas A (2005) Motif-based protein Detection of nonlinearity in multivariate cali-
sequence classification using neural networks. bration. Anal Chim Acta 376(2):153–163
J Comput Biol 12:64–82 24. Mallows CL (1986) Augmented partial residu-
17. Griffin WO, Razorilova S, Kitaev M et al als. Technometrics 28(4):313–319
(2010) An artificial neural network evaluation 25. Drapper NR, Smith H (1981) Applied regres-
of tuberculosis using genetic and physiological sion analysis, 2nd edn. New York, NY, Wiley
patient data. AIP Conference Proceedings, 26. Brereton RG (2003) Chemometrics: data anal-
Little Rock, 1229, p 49–53 ysis for the laboratory and chemical plant. John
18. Bengio Y (2009) Learning deep architectures Wiley & Sons Ltd., West Sussex
for AI. In: Foundations and trends in machine 27. Shelton RO (2000) A neural network develop-
learning, Now Publishers Inc. 2(1):1–127 ment tool: NETS, NASA’s Johnson Space center.
19. Cybenko GV (1989) Approximation by super- Open Channel Foundation, Publishing software
positions of a sigmoidal functions. Math from academic and research institutions. http://
Control Signal Syst 2:303–314 www.openchannelfoundation.org/projects/
20. Gasteiger J (1993) Neural networks for chemists: NETS/
an introduction. VCH Publishers, New York, NY 28. Tarasenko OJ, Kitaev M, Alisherov A et al
21. Lavine BK (2000) Chemometrics. Anal Chem (2000) Polymorphisms of HLA-A and -B
72(12):91R–97R genes in the Kyrgyz population. Anthrop Sci
22. Goicoechea HC, Collado MS, Satuf ML et al 4(108):293–304
(2002) Complementary use of partial least- 29. Rafei A, Pasha E, Jamshidi OR (2012)
squares and artificial neural networks for the Tuberculosis surveillance using hidden Markov
non-linear spectrophotometric analysis of model. Iran J Public Health 41(10):87–96
Chapter 18
Neural Networks and Fuzzy Clustering Methods

for Assessing the Efficacy of Microarray Based Intrinsic
Gene Signatures in Breast Cancer Classification
and the Character and Relations of Identified Subtypes
Abstract
In the classification of breast cancer subtypes using microarray data, hierarchical clustering is commonly
used. Although this form of clustering shows basic cluster patterns, more needs to be done to investigate
the accuracy of clusters as well as to extract meaningful cluster characteristics and their relations to increase
our confidence in their use in a clinical setting. In this study, an in-depth investigation of the efficacy of
three reported gene subsets in distinguishing breast cancer subtypes was performed using four advanced
computational intelligence methods—Self-Organizing Maps (SOM), Emergent Self-Organizing Maps
(ESOM), Fuzzy Clustering by Local Approximation of Memberships (FLAME), and Fuzzy C-means
(FCM)—each differing in the way they view data in terms of distance measures and fuzzy or crisp cluster-
ing. The gene subsets consisted of 71, 93, and 71 genes reported in the literature from three comprehen-
sive experimental studies for distinguishing Luminal (A and B), Basal, Normal breast-like, and HER2
subtypes. Given the costly procedures involved in clinical studies, the proposed 93-gene set can be used for
preliminary classification of breast cancer. Then, as a decision aid, SOM can be used to map the gene sig-
nature of a new patient to locate them with respect to all subtypes to get a comprehensive view of the
classification. These can be followed by a deeper investigation in the light of the observations made in this
study regarding overlapping subtypes. Results from the study could be used as the base for further refining
the gene signatures from later experiments and from new experiments designed to separate overlapping
clusters as well as to maximally separate all clusters.
Key words Breast cancer, Neural networks and fuzzy clustering, DNA microarray, Gene expression
signatures, Subtype classification, Gene clustering
1 Introduction
Breast cancer, a cancer in the breast tissues, is the second most

common cancer (after lung cancer) and the most common cancer
in women. Breast cancer is comprised of two types depending on
its original tissue: ductal carcinoma and lobular carcinoma. Breast
cancer molecular subtypes can be deciphered by using their distinct
285
286 Sandhya Samarasinghe and Amphun Chaiboonchoe
patterns of gene expression. There are at least five intrinsic subtypes:

luminal A (estrogen receptor [ER] and/or progesterone receptor
[PR] positive, HER2-), luminal B (ER and/or PR+, HER2+),
HER2 (ER and/or PR-, HER2+), basal-like (ER-, PR-, HER2-)
and normal breast-like. Recent understanding on the biology of
breast cancer can be found in a review by Carey [1].
Microarray technology provides an extensive source of gene
expression data that promise to pave the way for better prediction
and diagnosis of cancer and identification of key targets for drug
development. It allows simultaneous measurement of expression of
a large number of genes that is typically represented by an expres-
sion matrix in which the rows represent experiments or patients
and columns represent genes. Microarray data have been widely
used in breast cancer and recent reviews and details of microarray
based research on breast cancer can be found in [2–5]. Currently,
microarray research falls into two major themes:
●● Classification of breast cancer subtypes [6–8].
●● Development of prognostic gene signatures and clinical gene
markers to predict disease recurrence, response to therapy, sur-
vival outcome and drug sensitivity [9–17].
In classification of breast cancer subtypes using microarray
data, clustering is commonly used. This involves clustering gene
expression patterns (profiles, signatures) representing patients. As
unknown subtypes are typically sought, unsupervised clustering
methods are used where the expression profiles are divided into
highly similar groups without predefined group labels, based on a
similarity/dissimilarity or distance measure. The main clustering
method used so far in these studies is hierarchical clustering.
Although this reveals basic patterns of clusters, more needs to be
done to investigate the accuracy of clusters as well as to extract
meaningful cluster characteristics and relations from them to
increase confidence in the results for use in a clinical setting. To
this end, it is important to understand the features of the clustering
tools and how to select the appropriate tools in order to reveal
biological knowledge embedded in gene expression profiles in rela-
tion to molecular subtypes.
Breast cancer is a complex and heterogeneous disease. Many
studies have been conducted but the challenge still is to better
understand the biology of breast cancer. Microarrays have been used
to better understand molecular subtypes of breast cancer and this is
an ongoing research area: microarray data on breast cancer is now
available for public access and recent research has also been reviewed
by several groups [18–21]. There are many studies on breast cancer
gene signatures, including those that attempt to distinguish between
subtypes or find a prognostic gene signature. Some details on breast
cancer gene signatures from previous studies are provided in Table 1,
adapted from Weigelt, Peterse, and van’t Veer [22].
Breast Cancer Classification 287
Table 1
Microarray studies on prognostic gene-expression profiles in breast cancer (adapted from Weigelt,
Peterse, and van’t Veer [22])
Informative
Microarray type genes Outcomes References
cDNA 496 “intrinsic” Identified 4 subtypes [6]
genes :Luminal-like/ER+
gene, ERBB2+
overexpressed, Basal-
like, Normal breast-like
cDNA 456 “intrinsic” “Luminal A” tumors have [7]
genes a better outcome than
“luminal B” tumors.
Worst outcome is for
“basal-like” and
“ERBB2+” tumors
cDNA 534 intrinsic Tested results from [7] [8]
genes using independent
datasets
Oligonucleotide 70 genes “Identified a Good [12]
“70-gene signature” related to
Mamma low metastasis risk; a
Print” “poor signature”
associated with a high
metastasis risk.
Sensitivity: 91 %,
specificity: 73 %
Oligonucleotide 70 genes “Good signature” versus [36]
“poor signature.”
Sensitivity: 93 %,
specificity: 53 %
Oligonucleotide “metagenes” Accuracy: 90 % [37]
Oligonucleotide 76 genes “Good 76-gene [38]
signature” versus “poor
76 gene-signature.”
Sensitivity: 93 %,
specificity: 48 %
Oligonucleotide 442 genes Expression of “serum- [39, 40]
activated” signature
versus no expression.
Sensitivity: 91 %,
specificity: 29 %
(continued)
Table 1
(continued)
Informative
Microarray type genes Outcomes References
The “combined test set” data 306 genes Validated the breast cancer [27]
from Sorlie et al. [7, 8] (cDNA “intrinsic subtypes”
microarrays), van’t Veer et al. [12] by combining existing
(custom Agilent oligo microarrays) and data sets
Sotiriou et al. [28] (cDNA microarrays)
The “combined test set” data 50 genes Incorporated “intrinsic” [42]
from Sorlie et al. [7, 8] (cDNA subtype subtypes into risk model
microarrays), Hu et al. [27] predictor for breast cancer
and Perreard et al. [41] [GEO: “PAM50” prognosis and
GSE2607] (Agilent Human A1, Agilent prediction
Human A2, and custom oligonucleotide)
Oligonucleotide Affymetrix U133a 306 genes Identified relations [43]
microarrays: Wang and colleagues between breast
[GEO:GSE2034], Miller and colleagues subtypes, pathway
[GEO:GSE3494], and Pawitan and deregulation, and drug
associates [GEO:GSE1456] sensitivity.
We focus on breast cancer subtype classification in this chapter.

As shown in Table 1, Perou et al. [6] is the first group to focus on
distinguishing breast cancer subtypes, having carried out a series of
microarray experiments in 2000, 2001, and 2003, with subsequent
analysis based on hierarchical clustering. Their objective was to
identify genes characteristic of subtypes. From a long list of differ-
entially expressed genes characterizing the subtypes, they were able
to extract a small number of “intrinsic gene subsets” for each sub-
type. More recent studies tend to use the combined samples from
these previous studies and have resulted in a further reduced num-
ber of informative genes [11]. A review of the current situation of
breast cancer classification, in terms of classification efficacy, shows
that the search is still on to identify the most crucial “intrinsic
genes” that can be used at the clinical level to effectively classify
breast cancer subtypes which, in turn, can be incorporated into
cancer prognosis and prediction. As stated earlier, most of the pro-
posed intrinsic genes and subsets have been analyzed and identified
using hierarchical clustering. Therefore, this study attempts to
study the consistency, character, and relations of the subtypes based
on the intrinsic gene subsets reported in literature, when viewed
through the lens of different and more advanced clustering meth-
ods. The two main objectives in this study are as follows:
Objective 1: Study the effectiveness of the reported intrinsic gene
subsets in representing the breast cancer subtypes.
Objective 2: Test the efficacy of the gene subsets in classifying subtypes

and investigate the relationship/similarity between subtypes
In order to achieve these objectives, we study gene clustering
and patient clustering. We used the original “intrinsic” genes from
Sorlie et al. [7, 8] for gene clustering. We used four selected
clustering tools: Self-Organizing Maps (SOM) [23], Emergent

Self-Organizing Maps (ESOM) [24], Fuzzy clustering by Local
Approximations of MEmberships (FLAME) [25], and Fuzzy
C-Means (FCM) [26]. In patient clustering, we compare the
“intrinsic gene” sets from Sorlie et al. [8] and Hu et al. [27] using
three clustering methods: SOM, SOM clustering aided by the
Ward Method, and FCM.
2 Materials
2.1 Dataset There are three main datasets used in this study:
The first dataset was from Sorlie et al. [7]. Specifically, a
71-gene subset selected from 456 intrinsic genes as shown in
Fig. 1a: Group C—ERBB2+/HER2 (11 genes), Group D—Novel
unknown (12 genes), Group E—Basal-like (19 genes), Group F—
Normal breast-like (14 genes), and Group G—Luminal (15 genes).
These were derived from the raw data from a total of 85 cDNA
microarray experiments on tumor and normal breast tissues (71
ductal, five lobular, two ductal carcinoma in situ, three fibroade-
noma (benign tumor), and four normal breasts).
The dataset was retrieved from http://genome-www.stanford.
edu/breast_cancer/mopo_clinical/ or http://www.ncbi.nlm.nih.
gov/projects/geo/query/acc.cgi?acc=GSE3193.
The second dataset was from Sorlie et al. [8]. We used a
93-gene subset derived from 534 intrinsic genes (see Fig. 1b):
group C (ERBB2+/HER2 (11 genes)), group D (Luminal
Subtype B (28 genes)), group E (Basal-like) (24 genes), group F
(Normal breast-like) (5 genes), and group G (Luminal Subtype A)
(25 genes). These were derived from a total of 122 samples from
Norway/Stanford cohort which included the above mentioned 85
samples from the previous (2001) study used to classify cancer sub-
classes by average linkage hierarchical clustering.
The dataset was retrieved from http://genome-www.stanford.
edu/breast_cancer/robustness/ (also available at http://www.
ncbi.nlm.nih.gov/projects/geo/query/acc.cgi?acc=GSE4335).
In the above studies the raw data (cDNA) were processed and
genes were selected with a criterion of at least fourfold expression
from the median in at least three of the samples tested. Then
differentially expressed genes were used to find the intrinsic genes
by calculating the “within-pair variance” score and the “between-
subject variance.” The ratio between the two scores is called D and
intrinsic genes are those with a small value of D. This resulted in
a b c
290
ERBB2AA443351 TLK1 mucin 1, transmembrane

Steroidogenic acute regulatory protein related AA504615 TRAP100 DAZ associated protein 2
transcription elongation factor A SII-like 1
GRB7 H53702 PPARBP hypothetical protein FLJ21827
TGFB1-induced anti-apoptotic factor 1 AA446222 ERBB2 solute carrier family 9 sodium/hydrogen exchanger, isoform 3
TNF receptor-associated factor 4 AA598826 keratin 18
Flotillin 2 R72913
GRB7 programmed cell death 4 neoplastic transformation inhibitor
ERBB2 cyclin G2
Hypothetical protein FLJ10700 W81185
v-myb myeloblastosis viral oncogene homolog avian
SWI/SNF related, subfamily e, member 1 W51779 complement component 4A
fucosyltransferase 8 alpha 1,6 fucosyltransferase
trefoil factor 3 intestinal
basic transcription factor 3
X-box binding protein 1
estrogen receptor 1
GATA binding protein 3
solute carrier family 39 metal ion transporter, member 6
LPS-responsive vesicle trafficking, beach and anchor containing
rabaptin, RAB GTPase binding effector protein 1
J domain containing protein 1
glutathione S-transferase M3 brain
DKFZp564J157 protein
TNF receptor-associated factor 4

hypothetical protein FLJ10700
ATP-binding cassette, sub-family C CFTR/MRP, member 3
v-erb-b2
SRY sex-determining region Y-box 9 AA400464 growth factor receptor-bound protein 7
non-metastatic cells 2, protein NM23B expressed in
UDP-N-actyl-alpha-D-galactosamine H13688 non-metastatic cells 1, protein NM23A expressed in
ESTs W93120 CGI-48 protein
Cadherin 3, P-cadherin placental AA425217 ATP synthase, H+ transporting, mitochondrial F0 complex
Laminin, gamma 2 nicein AA677534 hypothetical protein FLJ13855
CXCL1 clathrin, heavy polypeptide Hc
Small inducible cytokine subfamily D R66139 H3 histone, family 3B H3.3B
ATDC AA055485 CDH3
Keratin 17 AA026642
interferon, gamma-inducible protein 30
Keratin 5 W72110 AHXA8 interferon induced transmembrane protein 1 9-27
Troponin I, skeletal, fast AA181334 KRT5 interferon-stimulated transcription factor 3, gamma 48kDa
Chitinase 3-like 2 AA668821 interferon, alpha-inducible protein clone IFI-6-16
Secretory protese inhibitor antileukoproteinase AA026264 TRIM29 interferon, alpha-inducible protein clone IFI-15K
KRT17 signal transducer and activator of transcription 1
Nuclear factor I/B W87528 interferon stimulated gene 20kDa
ESTs AI304356 MFGE8 chemokine C-X-C motif ligand 9
Transforming growth factor, beta 2 N48082 CX3CL1 caspase 1
Calpain-like protease AA457238 N-myc and STAT interactor
FZD7 v-yes-1 Yamaguchi sarcoma viral related oncogene
Dystrophin muscular dystrophy superoxide dismutase 2, mitochondrial
Fatty acid binding protein 7, brain W72051 CHI3L2
GRO1 oncogene, alpha W42723
B3GNT5
LRBA
Putative G protien-coupled receptor H50224

acyl-Coenzyme A dehydrogenase H95792 NAT1
estrogen receptor 1 AA291749 LIV-1
Trofoil factor 3 intestinal N74131
GATA-binding protein 3 H72474
HNF3A
GATA-binding protein 3 R31441 XBP1
X-box binding protein 1 W90128
Hepatocyte nuclear factor 3, alpha T74639 GATA3
Lymphoid nuclear protein related to AF4 H99588 ESR1
LIV-1 protein, estrogen regulated H29315
ESTs AA029948
PTP4A2
Hypothetical protein FLJ11280 N54608 RERG
N-acetyltranferase 1 arylamine R91802 SCUBE2
Myosin VI AA625890
Myosin VI AA030004
Fig. 1 Hierarchical clustering analysis from three previous studies [7, 8, 26]. The arrows indicated the same subtypes clustered by different gene sets. The red box
represents Basal-like subtype, the blue box represents Luminal A subtype and the pink box represents ERBB2+/HER2 subtype. (Names of the genes in some of
these subtypes can also be found in Table 4)
456 intrinsic genes from the first dataset and 534 intrinsic genes
from the second dataset. The authors then analyzed these genes by
hierarchical clustering, resulting in five distinct subtypes. Finally, a
small number of novel genes representing each subtype were
identified in these studies, as shown in the highlighted boxes in
Fig. 1a and b for 456 and 534 genes, respectively.
The last dataset was retrieved from Hu et al. [27]. This had a
71-gene subset derived from 306 intrinsic genes (see Fig. 1c):
Group C—Luminal (21 genes), Group D—HER2 (11 genes),
Group E—Interferon-regulated cluster containing STAT1 (12
genes), Group F—Basal-like (14 genes), and Group G—
Proliferation cluster (13 genes). A total of 311 tumors and 4 nor-
mal breast cancer samples was created by combining three existing
datasets of Sorlie et al. [7, 8], van’t Veer et al. [12], and Sotiriou
et al. [28]. The gene subsets containing 71 of the 306 intrinsic
genes are highlighted in boxes in Fig. 1c. The dataset was retrieved
from the UNC Microarray Database (http://www.ncbi.nlm.nih.
gov/projects/geo/query/acc.cgi?acc=GSE1992 or https://
genome.unc.edu/pubsup/breastTumor/).
In our study reported in this chapter, we explored these small
subsets of intrinsic genes separately, specifically, 71 and 93-gene
subsets from Sorlie et al. [7, 8] and 71-gene subset from Hu et al.
[27]. Some additional details of these subsets and the studies are
given in Table 2.
Table 2
Summary of data, intrinsic genes, and identified subtypes in studies from 2000 to 2006 [6–8, 27]
2003 (Sorlie
2000 (Perou et al.) 2001 (Sorlie et al.) et al.) 2006 (Hu et al.)
No. of 42 individuals (36 85 tissue samples 122 individuals 315 samples combined
samples ductal + 2 representing 84 (including 84 from Sorlie et al. [7, 8],
lobular + 1 ductal individuals (71 from 2001) van’t Veer et al. [12]
carcinoma in ductal + 5 lobular + 2 and Sotiriou et al. [28]
situ + 1 ductal carcinoma in
fibroadenoma + 3 situ + 3
normal breast) fibroadenoma + 4
normal breast)
Intrinsic 496 456 534 306
genes
Subtypes 1. Luminal-like / 1. Luminal 1. Luminal A 1. Luminal A
ER+ gene 2. HER2 2. Luminal B 2. Luminal B
2. HER2 3. Basal-like 3. HER2 3. HER2
3. Basal-like 4. Normal breast-like 4. Basal-like 4. Basal-like
4. Normal breast-like 5. A novel unknown 5. Normal 5. Normal breast-like and
cluster breast-like 2 new clusters:
Proliferation,
Interferon-reg.
3 Neural Networks and Fuzzy Clustering Methods
Four methods were used in this study: Self-Organizing Maps

(SOM), Emergent Self-Organizing Maps (ESOM), Fuzzy Linear
Approximation of Membership (FLAME), and Fuzzy c means
(FCM). A summary of these methods is given in the subsections
that follow:
3.1 Self-Organizing An artificial neural network (ANN) is a collection of intercon-

Feature Map (SOM) nected neurons that incrementally learn from their environment to
capture essential linear or nonlinear trends in complex data, so that
it provides reliable predictions for new situations containing even
noisy and partial information [29]. There are various ANNs that
can be constructed and used for various purposes. This research
study uses Self-Organizing Feature Map (SOM). Self-organizing
maps are unsupervised neural networks that preserve the exact
topology of data space on a two-dimensional grid of neurons. The
components of SOMs are neurons that are normally arranged in a
two-dimensional hexagonal or rectangular grid, as shown in Fig. 2.
The input layer represents the input variables x1, x2, …, xn for the
case of n inputs and each neuron (circles in Fig. 2) is associated
with an n-dimensional weight vector wi = [w1, w2, …, wn]. The
competition among the neurons is the mechanism responsible for
self-organization in the brain, a fact which is utilized in developing
SOM networks. Through a training process involving such compe-
tition to iteratively modify weights, inputs (e.g., gene expression
vectors) are mapped from high dimensions to this two-dimensional
map space, where cluster structure as well as proximity of clusters
to each other can be visualized.
Training starts with assigning to neurons random weight vectors
with the same dimensions as the input vectors; Euclidean distance
(or any other distance measure) is then used to calculate the distance
between an input vector and each of the weight vectors, as shown in
Eq. 1. The neuron with the weight vector closest to the input vector
(i.e., with the smallest distance) is declared the winner.
∑(x )
2
Euclidean Distance = d j = x − w j = i − wij (1)
i =1

The objective of learning in SOM networks is to accurately project
the input data onto the map neurons. Over repeated exposure to
inputs during learning, neurons learn to respond to inputs in such
a way that neurons that are closer on the SOM respond to inputs
that are closer in the actual data space. This feature, known as
topology preservation, is achieved by moving not only the winner
neuron but also its neighbor neurons towards the input in order to
minimize the distance between these neurons and the input. In
x1 x2 xn
Fig. 2 Self-organizing feature map
this process, the winner moves by a larger amount than its neigh-
bors according to a Neighbor Strength (NS) function.
For each neuron j ∈ Nc, where Nc is the neighborhood sur-
rounding the winner neuron c, learning involves adjusting the new
weight as follows:
w new
j = w old
j + ε (t ) NS ( d ,t )  xi − w old
j 
 (2)

where wjnew is new weight, wold is the previous weight, ɛ is learning
rate, xi is the ith input vector, and NS is a neighborhood function
defining the strength of weight adjustment of neighbors with respect
to the winner at iteration t and can take the form of a linear or nonlin-
ear function of distance d from the winner. The learning rate ε(t) is
user-defined and can take the form of a linear, e xponential, etc. func-
tion. Starting with a larger value, ε(t) decreases with iterations. NS(d,t)
also can be linear, exponential, Gaussian, etc. and shrinks with itera-
tions so that an initially large neighborhood size gets reduced to just
the winner, or the winner and its immediate neighbors at the end of
training. Training stops when there is no acceptable difference in
weight adjustment in successive iterations. In this study, we used the
Neural Networks Toolbox (Matlab) for training the SOM.
SOM training results in a map where inputs that are closer in
the data space are projected onto neurons that are closer on the
SOM. Therefore, any large gaps between data in the data space are
shown as large distances between corresponding neurons on the
map, thereby revealing on the map potential clusters and cluster
boundaries in the actual data. With an appropriate clustering tool,
such as Ward Clustering, neurons that form clusters on the map
can be grouped to form classes for classification purposes. An
unknown input vector then will belong to the class that it is closest
to, i.e., the class with the shortest distance between its mean vector
and the input vector. This provides a crisp classification for the
input vector—one input vector belongs to one class.
In a trained SOM, each neuron represents the center of gravity
of a number of input vectors. SOM results can be visualized in vari-
ous ways. In one form, a U-matrix (Unified distance matrix) the
average distance between a neuron and its neighbors is shown with
colors on the map. This can show distinct subtypes or gene clusters
depending on whether patient or gene clustering is done. Another

visualization scheme shows the spread of values of each input vari-
able (e.g., expression of a gene across all patients or expression of
all genes in a patient depending on whether genes or patients are
clustered) in a separate map pane. In order to quantify clusters,
trained SOM neurons (weight vectors) are subjected to a cluster-
ing algorithm. In this study, the Ward method, an efficient cluster-
ing method suitable for small datasets, was used to find clusters of
genes and cancer subtypes on the SOM using an in-house devel-
oped algorithm implementing Ward clustering.
Ward clustering [30] is an agglomerative clustering method
invented in 1963. For incrementally forming clusters, the method
calculates the distance between data and cluster centers for all
potential clusters by analyzing the variance amongst them.
Specifically, it aims at minimizing the sum of square distance
between all data points in a cluster and the centroid of the cluster
formed as a result of joining two clusters. The two clusters that
produce the least square distance are ultimately joined in each step
of the process. The distance is calculated as:
d ward =
(n r × n s ) xr − xs
2
(3)
(n r + n s )
where r and s symbolize two clusters and nr and ns are the number
of data points in those clusters. ||xr – xs|| gives the magnitude of
the distance between two cluster centers calculated using
Euclidean norm. The Ward distance increases with the increase in
the number of data points. The two clusters with minimum dward
are joined at each step of this process. The ward index gives the
likelihood of each cluster. The separation of clusters depends
upon the value of the index. The higher the index, the more dis-
tinctive are the clusters.
1 d − dt −1 1 ∆dt
Ward Index = × t = × (4)
NC dt −1 − dt − 2 NC ∆dt −1

where ∆dt is the distance between the centers of two clusters to be
merged and ∆dt−1 is the distance between the centers of the clusters
merged previously and prior to that step. NC is the number of
clusters remaining. The method is very effective when the dataset
is not too large. Therefore, it is quite suitable for use on a trained
SOM to cluster map neurons. In this case, it uses the neuron
weights generated by the trained SOM as input vectors.
3.2 Emergent ESOM is an extension of SOM which allows a map to grow in size from
Self-Organizing the initial map. ESOM is arranged in a toroidal grid using a larger num-
Maps (ESOM) ber of neurons than a typical SOM [31]. An ESOM-map can be visual-
ized in three forms: U-Matrix (distance-based), P-Matrix (density-based)
and U*-Matrix (distance and density based). The U-Matrix, which

indicates the average distance of each neuron to its nearest neighbors is
the same as the U-Matrix in SOM, so this study uses the P-Matrix in
order to analyze density of gene expression data. The P-Matrix contains
entries denoting density of neurons in the neighborhood of a neuron.
A neuron with a large P is located in a dense data region while a small p
indicates sparse data regions in the data space. For neuron n, the density
P(n) in the data space X can be defined as:
P (n ) = p (w (n ) , X ) , (5)

where p(x,X) denotes an empirical density estimation at point x
(i.e., w(n) which is the weight vector of neuron n) in the data space
X and ESOM uses Pareto Density Estimation (PDE) [31], an
optimal information kernel density estimation.
Data can be analyzed by using publicly available ESOM soft-
ware (Databionics ESOM Tool) developed by Ultsch and Morchen
[31]. This software is available to download from http://
databionic-esom.sourceforge.net/.
3.3 Fuzzy Clustering The Fuzzy clustering by Local Approximation of MEmbership

by Local (FLAME) algorithm starts by calculating similarities of gene
Approximation expression patterns using Pearson correlation, and then creating a
of MEmbership connected graph of all K-Nearest Neighbors (KNN) of each
(FLAME) expression vector. For each expression pattern, density, which indi-
cates the number of neighbors within a specified distance, is calcu-
lated to classify it into one of three groups: (1) cluster supporting
object (CSO) which has higher density than their neighbors, (2)
outlier, which has lower density than its neighbors and below a
predefined threshold, and (3) the rest. In the next step, the Local
Approximation of fuzzy membership (initialised to random values)
of the three groups is determined and each object in group 3 (rest)
is updated by a linear combination of the fuzzy memberships of its
nearest neighbors. The degree of membership of vector x in cluster
i is pi(x) and is given by:
x : p ( x ) = ( p1 ( x ) , p2 ( x ) ,…, pM ( x ) ) , (6)

M
where 0 ≤ pi (x ) ≤ 1; ∑ pi (x ) = 1 and M is the total number of clus-
i =1
ters: M = |Xcso| + 1, where Xcso is the set of cluster supporting objects
with Local Maximum Density, and 1 represents the outlier cluster.
The p(x) is represented by a membership vector of each object
x and is updated by the weighted summation of membership vec-
tors of all its nearest neighbors, as in Eq. 7.
p t +1 ( x ) ≈ ∑ wxy p t ( y )
y ∈KNN ( x )
(7)

where pt(x) is a fuzzy membership vector of object x at iteration t

and wxy are the coefficients reflecting relative proximities of nearest
neighbours such that the sum of the coefficients over the KNN(x)
is equal to 1 and KNN is the k nearest neighbors of x.
Basically, each iteration process attempts to reduce the Local
(Neighborhood) Approximation Error E [24] which is the differ-
ence between the approximation of membership vectors in the cur-
rent and previous iterations, defined by:
2
E ({ p } ) = ∑ p ( x ) − ∑ wxy p ( y ) (8)
∀x y ∈KNN ( x )

Finally, based on fuzzy memberships the clusters can be repre-
sented in two ways: one gene to one cluster or one gene to multi-
ple clusters. FLAME results are given as line plots showing three
patterns: (1) each cluster which includes CSO and the rest, (2) all
outliers, and (3) all CSOs. The most important is the type (1)
which shows expression patterns for each cluster in a separate line
plot. FLAME is integrated with Gene Expression Data Analysis
Studio (GEDAS) (http://sourceforge.net/projects/gedas).
3.4 Fuzzy FCM is the most widely used fuzzy clustering algorithm; it was
C-Means (FCM) developed by Dunn in 1973 [32] and improved by Bezdek in 1981
[33]. It is based on the minimization of the following objective
function:
c c
J = ∑J i = ∑ ∑ u d (x m
ki k − ci ) , 1 ≤ m < ∞ (9)
j =1 i =1 k , x k ∈ci

where ci is the centroid of cluster i, d(xk − ci) is the distance between
ith centroid (ci) and kth input vector xk, c is the total number of clus-
ters, m is a fuzziness parameter which can be any real number greater
than 1, and uki is the degree of membership of xk in cluster i.
FCM is a clustering method that allows data to belong to more
than one cluster with a degree membership uij. The FCM algorithm
starts with a predefined number of clusters where each datum has
an equal degree of belonging to every cluster. Fuzzy partitioning is
carried out through an iterative optimization of the objective func-
tion shown above, with updates of membership uij and cluster cen-
ters cj. The iteration stops when a termination criterion is met. FCM
is integrated with Gene Expression Data Analysis Studio (GEDAS).
Each of the selected methods allows the user to adjust param-
eters. We set parameters for each method as follows: SOM based
on correlation distance and batch learning was conducted on
Matlab (http://www.mathworks.com/) Neural Networks toolbox
followed by an in-house developed algorithm implementing Ward
clustering to define clusters. ESOM based on correlation distance
used online (example-by-example) learning while growing map

size to 50 × 82 in 20 batch runs; linear functions were used for
both shrinking the neighborhood size and learning rate. FLAME
used Pearson or Squared Pearson Correlation with 10K-Nearest
Neighbors with no threshold. FCM used Pearson or Squared
Pearson Correlation distance with fuzziness parameter (m) of 1.2.
This fuzziness parameter was calculated by using an empirical
equation from Dembele and Kastner [34]. They proposed that the
value of m should be lower than or equal to 2 in order to get gene
members with strong relationships to clusters. Ozkan and Turksen
[35] also mention that level of fuzziness is essential and it can affect
FCM membership values.
4 Results and Discussion
4.1 Gene Clustering In this section, the results from clustering the small subsets of 71
Based on the 71 and 93 genes (extracted from the two intrinsic gene sets of 456
and 93 Genes by SOM, and 534, respectively) using SOM, ESOM, FLAME, and FCM are
ESOM, FCM, presented, highlighting how these two small gene subsets support
and FLAME the identified cancer subtypes by assessing their clustering patterns
when subjected to these clustering algorithms. As highlighted
earlier, the 71-gene subset from the 456 genes represented four
subtypes: HER2 (refers to ERBB2+/HER2), Basal-like, Normal
(i.e., Normal breast-like) and Luminal. The 93-gene subset from
the 534 genes represented five subtypes: HER2, Basal-like,
Normal, Luminal A, and Luminal B.
SOM was developed and the map was clustered using the Ward
method to find optimum clusters. Results of clusters and corre-
sponding Ward index are shown in Fig. 3. Maximum Ward likeli-
hood index for both datasets as evidenced by the largest Ward
likelihood index indicates that the optimum number of clusters in
each case is three. Table 3 shows details of each SOM-based gene
cluster in terms of the number of genes from the small gene subsets
of 71 or 93 represented in each cluster against the number of genes
in the original subtypes.
The fact that both gene sets resulted in three clusters indicates
the strong likelihood that not all five subtypes are distinct. The
members in each SOM cluster for the first dataset (Table 3a) rep-
resent two or more subtypes. Cluster 1 is divided between Luminal
and Normal breast-like (hereafter referred to as either Normal-like
or Normal), which implies that Luminal and Normal-like have
some similarity in gene expression patterns. In contrast, the
93-gene subset extracted from the 546 “intrinsic” genes shows
clear separation of Luminal A into cluster 1 (Table 3b). However,
Table 3b indicates a weak similarity between Basal-like and Normal-
like in its cluster 2 which is predominantly Basal-like. Similarly,
cluster 2 in Table 3a is predominantly Basal (13/19) with some
Normal-like (3/14) cases and one of HER2 (1/11) subtype
a b
6 6
5 5
4 4
3 3
2 2
1 1
0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60 70 80 90 100
Fig. 3 Ward likelihood index for various number of clusters (a) for 71 genes from 456 “intrinsic” gene set (b)
for 93 genes from 546 “intrinsic” gene set ((x-axis—number of clusters and y-axis—Ward likelihood index).
The larger the likelihood index, more likely that number of clusters
Table 3
Proportion of genes belonging to the original subtypes found in each SOM cluster (numerator is the
number of genes belonging to a particular subtype found in an SOM cluster and the denominator is
the total number of genes belonging to that subtype)
Number of genes in original subtype/number of genes in SOM cluster 1 2 3

(a) Clustering based on the 71-gene subset extracted from the 456 “intrinsic” genes
HER2 (11 genes) 1/11 10/11
Basal-like (19 genes) 13/19 6/19
Normal (14 genes) 11/14 3/14
Luminal (15 genes) 14/15 1/15
(b) Clustering based on the 93-gene subset extracted from the 534 “intrinsic” genes
HER2 (11 genes) 11/11
Basal-like (24 genes) 18/24 6/24
Normal (5 genes) 4/5 1/5
Luminal A (28 genes) 28/28
Luminal B (25 genes) 25/25
s upporting not only the possible similarity between Basal-like and

Normal-like but also showing the potential to overlap with the
HER2 subtype. Cluster 3 in Table 3a is dominated by HER2
(10/11) with a moderate presence of Basal-like (6/19) and weak
presence of Luminal (1/5). Cluster 3 in Table 3b also has all HER2
genes (11/11) but it also has all the Luminal B gene activity pat-
terns indicating a strong similarity between HER2 and Luminal
B. The rest in this cluster are from Basal-like and Normal-like

whose similarity was already revealed in cluster 2 of Table 3a, b.
Overall, Luminal A appears distinct from Luminal B and others,
Basal and Normal-like show overlap, and HER 2 shares similarities
with Basal-like and Luminal B.
ESOM results, given in Fig. 4, show information on the gene
density for the two intrinsic gene sets. The red color represents
dense areas while the lighter color represents sparse areas. The dots
indicate neurons on the map. In this figure, the toroidal shape of
the map has been converted to a rectangular shape for visualization
purposes but one can visualize the toroid by bending the corners
appropriately. This map also shows three dense areas (red to yellow
areas in the corners and within) for both gene sets. A further clus-
tering step was attempted but clusters seemed unstable in multiple
trials and therefore, the ESOM results are shown here mainly to
present a visualization of the characteristics of data density, not for
clustering purposes. However, we use the results to highlight
where the genes for three subtypes tend to cluster by investigating
the best match ESOM vector to each gene expression vector
(Fig. 5) for the case of 93-gene subset from the 534 intrinsic genes.
For example, Basal-like cluster has all its genes in the same location
while HER2 has all but SMARCE1 in the same location. Normal-
like cluster is close to the Basal-like and has all its genes in the same
location except for KIAA0857 and MEN1.
The last two clustering methods (FLAME and FCM) provide
cluster information which can be used to compare with the original
clusters and with SOM clusters. In doing so, the focus is on inves-
tigating the efficacy of the gene subsets in distinguishing the sub-
types as found by the original authors by analyzing if each subset is
clustered separately. We present in Table 4 the results of clustering
tumor subtypes using the two sets of 71 and 93 genes by the three
methods (SOM, FLAME, and FCM) and compare with the
clusters from the original study. Recall that SOM produced three
optimum clusters as shown in Table 3. FLAME and FCM used squared
Pearson correlation. Squared Pearson captures the similarity between
profiles in terms of trends irrespective of whether upregulated or
downregulated. These can help identify co-regulated genes which
may be either upregulated or downregulated in a related fashion.
FLAME finds the optimum number of clusters by itself and
produced four clusters for the 71 genes, similar to the number of
original subtypes, as shown in Table 4. FCM requires the number
of clusters to be predefined so the same number of clusters as the
original subtypes was given to FCM in order to compare their
results. In Table 4, cluster results are presented according to the
number label given by the computational method to each specific
cluster followed by a number within brackets that indicates the
number of genes in the original subset for each subtype found in
the computed clusters. The question we ask is: Do the clusters
ESOM: 71 genes from 456 intrinsic genes
45
40
35
30
25
20
15
10
5
0
50
45
40 90
35
30 80
25 70
20 60
50
15 40
10 30
5 20
0 10
0
ESOM: 93 genes from 534 intrinsic genes
70
60
50
40
30
20
10
0
50
45
40 90
35 80
30 70
25 60
20 50
15 40
10 30
20
5 10
0 0
Fig. 4 ESOM: P-Matrix (density-based) and 3D view of P-Matrix for the two datasets. White dots represent best
match (closest) vector to each data point. Red indicates higher density and blue indicates smaller density of
input vectors
SMARCE1
KIAA0857, MEN1
Basal-like
HER2
Normal-like
Fig. 5 Mapping of three cancer subtypes (found by original authors from the 93-gene subset) onto ESOM
s upport the gene sets identified as representing the cancer subtypes?

Results indicate that for the 71-gene set, FCM put all four subsets
of genes (subtypes) into four distinct clusters. FLAME put all but
Basal-like into distinct clusters. Even in the case of Basal-like, most
(68.4 %) are in one cluster (Cluster 5) and the rest shared by
Luminal (26.34 %) and Normal-like (5.26 %), further supporting
Table 4
Clustering results from the three methods: SOM, FLAME, and FCM using 71 and 93 genes (number in
brackets refers to number of genes from the original subtypes from Sorlie et al. [7, 8])
Subtype Subtype genes (original study) SOM FLAME FCM

71 genes extracted from 456 intrinsic genes Sorlie et al. [7]
HER2 ERBB2, ESTs T57034, KIAA0130, Cluster 2 (1) + Cluster 3 (11) Cluster 3 (11)
(11) ERBB2, steroidogenic, GRB7, cluster 3
TGFB1, TNF, flotillin 2, (10)
FLJ10700, SWI/SNF
Basal-like SRY, UDP-N, ESTs W93120, CDH Cluster 2 (13) + Cluster 0 (5) + Cluster 0 (19)
(19) 3, laminin, small inducible cluster 3 (6) cluster 2 (1) +
cytokine subfamily, ATDC, KRT cluster 5 (13)
17, KRT 5, troponin I, CHI3L2,
SLPI, nuclear factor I/B, ESTs
AI304356, transforming growth
factor beta 2, calpain-like
protease, dystrophin, fatty acid
binding protein 7, GRO1
Normal CD36, CD36, glutathione peroxidase Cluster 1 (11) + Cluster 2 (14) Cluster 5 (14)
(14) 3, glycerol-3-phosphate cluster 2 (3)
dehydrogenase 1, lipoprotein
lipase, ESTs T62068, four and a
half LIM domains 1, retinol-
binding protein 4, amine oxidase,
integrin, alpha 7, alcohol
dehydrogenase 2 (class I), MY047
protein, aquaporin 7, 30 kDa
protein
Luminal Putative G, acyl-Coenzyme A, Cluster 1 (14) + Cluster 0 (15) Cluster 2 (15)
(15) ERS1, TFF3, GATA3, GATA3, cluster 3 (1)
XBPI, HNF3A, lymphoid nuclear
protein, LIV1, ESTs AA029948,
FLJ 11280, N-acetyltransferase,
Myosin VI, Myosin VI
93 genes extracted from 534 intrinsic genes Sorlie et al. [8]
HER2+ FLJ10700, FLOT2, SMARCE1, Cluster 3 (11) Cluster 2 (11) Cluster 0 (11)
(11) TLK1, TRAP100, GRB7,
PPARBP, ERBB2, ERBB2,
Homo sapiens mRNA,TBPL1
Basal-like CXCL1, CDH3, ANXA8, KRT5, Cluster 2 (18) + Cluster 1 (24) Cluster 1
(24) TRIM29, KRT17, MFGE8, cluster 3 (6) (8) + cluster 2
CX3CL1, FZD7, CHI3L2, (15) + cluster
B3GNT5, EXT2, FLJ10697, FLJ 3 (1)
31360, FLJ 14761, SLPI, EST,
TONDU, GABRP, ZDHHC5,
FLJ 11796, SLC5A6,GABRP,
FLJ 11796
(continued)
Table 4
(continued)
Subtype Subtype genes (original study) SOM FLAME FCM

Normal KIAA0857, MEN1, PIK3R1, Cluster 2 (4) + Cluster 1 (3) + Cluster 2 (5)
(5) AKR1C1, FACL2 cluster 3 (1) outlier (2)
Luminal LRBA, Homo sapiens mRNA, LIV-1, Cluster 1 (28) Cluster 3 (28) Cluster 2
subtype HNF3A, XBP1, GATA3, ESR1, (7) + cluster 3
A (28) PTP4A2, RERG, ACADSB, (1) + cluster 4
GATA3, VAV3, KIAA1243, (20)
LOC51313, DKFZp, KIA0876,
FLJ 10980, TLE3, MGC 22588,
CEGP1, FBP1, MGC 27171, HSD
17B4, Homo sapiens Mrna, Homo
sapiens mRNA, NAT1, Homo
sapiens, FLT1
Luminal ABCD3, ADSL, BTG3, ATP5G1, Cluster 3 (25) Cluster 4 (25) Cluster 3
subtype D123, PRNPIP, EBNA1BP2, (24) + cluster
B (25) NSEP1, GGH, LC27, PRDX4, 4 (1)
HSPC163, GARS, FLJ12442,
TMSB10, KDELR2, KIAA1691,
NRBF-2, CCNE1, CALU,
LANP, POLR2F, SQLE,
S100A10
the evidence for some similarity between these three subtypes.

As discussed before, SOM (i.e., SOM/Ward) produced only three
clusters and its performance overall is weaker than that for FLAME
and FCM. The SOM puts 93 % Luminal and 78 % Normal-like in
one cluster, Normal, Basal-like and HER2 in one cluster, and
HER2, Basal-like, and Luminal in another cluster, thereby splitting
the original subtype genes as shown clearly in Table 3a. Closer
inspection reveals that SOM cluster 3 is predominantly HER2 with
91 % (10/11) HER2 genes and SOM cluster 2 is predominantly
Basal-like with 68 % (13/19) Basal-like genes in it (Table 3a). Thus,
although SOM clusters correspond with some subtypes, they also
share genes from other subtypes; especially, in the case of Normal,
Luminal, and Basal-like. The largest split was for Basal and FLAME
also has difficulty demarcating the Basal-like subtype.
To complement results for the 93-gene subset in Table 4 and to
clarify the results better, the same results are produced in summary
form in Table 5 to shed more light on the efficacy of this 93-gene
set extracted from the 534 intrinsic genes (note that these 534
genes were found from an extension of the study that produced the
456 genes) in identifying the five subtypes. Specifically, Table 5
indicates the percentage of the genes defining the original subtypes
found in each cluster from SOM, FLAME, and FCM. Results in
Table 5b show that FLAME produced four optimum clusters and
all but Normal-like and Basal-like genes were put into distinct
clusters. Specifically, some (60 %) Normal-like genes were clustered
with Basal-like and the rest of the Normal-like genes were consid-
ered as outliers (see Table 5b). Note though that the Normal-like
subset now has only five genes and this may have affected the results.
SOM (Table 5a) dedicated one cluster (cluster 1) to Luminal A;
and another (cluster 2) to Basal-like and Normal-like genes but put
roughly 25 % of genes in these subtypes in cluster 3 where all
Luminal B and HER2 genes have also been placed. Thus cluster 3
represents four subtypes with the majority (58 %) of the cluster
being genes from Luminal B followed by HER2 (26 %), Basal-like
(14 %) and only 2 % of Normal-like genes. Overall, FLAME shows
more definite support for original subtypes than SOM with Lum A,
Lum B, and HER2 in separate clusters and Basal and Normal-like
sharing one cluster. This shows that the FLAME’s choice of four
clusters as opposed to the original authors’ five subtypes is based on
Table 5
Percentage of genes from each subtype found in each cluster of (a) SOM/Ward, (b) FLAME, and (c)
FCM based on the 93-gene subset
Cluster 0 Cluster 1 Cluster 2 Cluster 3 Cluster 4

(%) (%) (%) (%) (%)
(a) SOM/Ward
LumA (28) 100
LumB (25) 100
HER2 (11) 100
Basal (24) 75 25
Normal (5) 80 20
(b) FLAME
LumA (28) 100
LumB (25) 100
HER2 (11) 100
Basal (24) 100
Normal (5) (outlier 40 %) 60
(c) FCM
LumA (28) 25 3.5 71.5
LumB (25) 96 4
HER2 (11) 100
Basal (24) 33.34 62.5 4.16
Normal (5) 100
the similarity between Basal-like and Normal-like subtypes that was

persistently shown across all results in this study.
FCM was used with five clusters to match the original number
of subtypes. It puts two gene subsets into distinct clusters (all
HER2 genes in cluster 0 and almost all Luminal B genes in cluster
3 and over 70 % of Luminal A genes in cluster 4 (Table 5c)).
However, it puts all Normal-like genes with the majority of Basal-
like genes in cluster 2 which also contain 25 % of Luminal A genes
(Table 5c). This again indicates some overlap between Basal-like,
Normal-like, and Luminal A subtypes. Interestingly, some 33 % of
Basal-like genes have gone into a separate cluster (cluster 1). The
FCM summary results in Table 5c show complete support for
HER2 subtype and strong support for LumB subtype with most of
the latter’s genes in cluster 3. All other subtypes are spread across
one or more clusters (e.g., LumA and Basal-like). Overall results
from both analyses (71 and 93 genes) show FLAME with strong
support for the identified subtypes followed by FCM. Support
from SOM is weak in both cases.
Similarity or distance measure used can affect the gene mem-
bership in each cluster. Therefore, selecting clustering tools with a
priori knowledge of the expected characteristics of genes will
enhance the confidence in and accuracy of results. To assess the
influence of the distance measure used, we conducted another
introspective study with only the 93-gene subset extracted from
the 534 genes found by Sorlie et al. [8]. Specifically, FLAME and
FCM are used here with Pearson or squared Pearson correlation as
the distance measure and the results are shown in Table 6.
FLAME does not need a predefined number of clusters and
identified only four clusters from both distance measures. FCM
needs to predefine the cluster number, so two options are studied
where five is given according to known subtypes from the original
authors’ work and four is given in order to compare with FLAME
clusters. Entries in this table indicate the cluster number (1–4 or
0–3) followed by two numbers within brackets: the number of
genes in the original authors’ subtype found in this cluster (numer-
ator) and the total number of genes in the cluster (denominator).
Table 6 shows that the distance measure and the defined number
of clusters (for FCM) have an effect on how genes are clustered.
Pearson gives better cluster results for both FLAME and FCM
(four clusters) than squared Pearson correlation.
Four clusters from FLAME and FCM based on Pearson cor-
relation are the most similar to the original group; an exception is
that both FLAME and FCM define Basal-like gene expression pat-
terns similar to those for Normal-like subtype (cluster 1 for FLAME
and cluster 2 for FCM), an observation previously made from our
SOM, FLAME, and FCM results. In fact, 4- (as well as 5-) cluster
FCM based on Pearson correlation indicated the similarity between
these two subtypes (cluster 2 contains all five Normal-like and
most or a large number of Basal-like genes). These agree with the
Table 6
Comparison of clustering results from FLAME and FCM using only the 93-gene subset from the 534 intrinsic genes with Pearson and Squared Pearson
correlation coefficient as the distance measure
FLAME FCM
Squared Pearson 5
Subtype Pearson 4 clusters Squared Pearson 4 clusters Pearson 5 clusters Pearson 4 clusters clusters
HER2 (11) Cluster 2 (11/11) Cluster 2 (11/11) Cluster 0 (11/11) Cluster 0 (11/11) Cluster 0 (11/11)
Basal-like (24) Cluster 1 (24/27) Cluster 0 (13/46) + Cluster 1 (8/8) + Cluster 2 (23/28) + Cluster 2 (23/27) +
cluster 3 (11/12) cluster 2 (15/27) + cluster 3 (1/26) cluster 4 (1/21)
cluster 3 (1/26)
Normal (5) Cluster 1 (3/27) + Cluster 0 (3/46) + Cluster 2 (5/27) Cluster 2 (5/28) Cluster 2 (23/27) +
outlier (2/2) cluster 3 (1/12) + Cluster 4 (1/21)
outlier (1/1)
Luminal subtype A (28) Cluster 3 (28/28) Cluster 0 (28/46) Cluster 2 (7/27) + Cluster 1 (28/28) Cluster 1 (8/9) +
cluster 3 (1/26) + cluster 4
cluster 4 (20/21) (20/21)
Luminal subtype B (25) Cluster 4 (25/25) Cluster 0(2/46) + cluster 1 Cluster 3 (24/26) + Cluster 3 (25/26) Cluster 3 (25/25)
(23/23) cluster 4 (1/21)
The numbers within brackets are genes from the original author’s (Sorlie et al. [8]) subtype found in each cluster (numerator) and the total number of genes in that cluster
(denominator)
Breast Cancer Classification
305
results from hierarchical clustering where Normal and Basal-like

are close to each other with similar gene expression patterns
(Fig. 1). FLAME identified that two genes, KIAA0857 and
PIK3R1, were distinct from the others in the Normal-like cluster
and classified them as outliers In fact, ESOM also put KIAA0857
separately from the rest of the genes in the Normal-like cluster
(Fig. 5). Both FLAME and FCM with four clusters clearly sepa-
rates Luminal A and Luminal B into distinct clusters; but if pressed
for five clusters, FCM puts some Luminal A genes with Basal-like
and one with Luminal B (see clusters 2 and 3 in Luminal A) and
splits Basal-like genes even further than before. Whether it is four
or five clusters, HER2 genes make a distinct cluster, indicating the
efficacy of these genes in characterizing this subtype.
Comparing FCM 5 clusters based on Pearson correlation and
Squared Pearson correlation, HER2 cluster is still intact, Luminal
B separates into a distinct cluster (marginal improvement), and
Basal-like cluster has become clearly more distinct separating very
much into one cluster (a large improvement). Luminal A cluster
has become very slightly more distinct, and Normal-like has slightly
deteriorated. With FLAME, squared Pearson correlation has dete-
riorated the clustering results. Thus, overall, four clusters based on
Pearson correlation seem to affirm the efficacy of the gene subsets
in distinguishing breast cancer subtypes with FLAME producing
more accurate clustering outcomes.
4.2 Classification Another useful investigation that can shed light on the selected gene
of Subtypes: subsets is clustering patients. Therefore, we conducted an SOM and
Clustering Patients FCM cluster analysis of patients based on the 93-gene subset extracted
Using SOM and FCM from the 534 intrinsic genes [8] and the 71 subset of genes extracted
from the 306 intrinsic genes [27] (see Table 2, last column). The lat-
ter set has been proposed from a combined dataset from three sources
including the one that produced the above 534 intrinsic genes. The
results from the previous section show that each clustering method
produces one or more clusters that contain genes from one or more
subtypes. Fuzzy clustering is a soft clustering approach that allows a
measure of the strength of membership of an item in each cluster and
we utilize this aspect of FCM more in this patient clustering.
4.2.1 SOM Clustering Gene expression vectors, each containing expression results for all
of Patients 93 genes (from the 534 intrinsic genes) for each of the 79 patients
were clustered using SOM and results are shown in Fig. 6. The first
map Fig. 6a (left) indicates the number of patients (hits) repre-
sented by each neuron, and Fig. 6a (right) shows the average dis-
tance between neurons where lighter color indicates smaller
distances and darker color indicates larger distances between neu-
rons. Such a distance map allows visualization of potential clusters.
Dark regions indicating larger gaps appear at least in three places
and these separate the samples into at least three groups.
Fig. 6 (a) Projection of all patients (subtypes) on to the trained SOM based on the 93-gene subset (left ), and
U-matrix (right ). (b) SOM sample hits (patients) for each “intrinsic” subtype according to the 93-gene subset
from the original study (79 patients in total)
In the last five panels (Fig. 6b), the actual expression pattern

for patients from each original subtype is projected onto the trained
map. A powerful feature of SOM is topology preservation where
data (expression patterns; in this case patients) that are biologically
closer are organized closer in the map allowing one to see not only
clusters but also biological similarity between different clusters
(subtypes). The SOM panels for the five subtypes show that most
patients of each subtype are organized in the same region; for
example, Normal-like patients are located at the bottom left-hand
corner of the Map. Moreover, Normal-like is located opposite to
LumB; Basal-like is located opposite to Luminal A; and HER2 is
closer to Normal-like and Basal-like in that these three subtypes
together occupy a triangular region in the bottom right part of the
map. LumA and LumB occupy the top left triangle on the map.
This corresponds to classification of breast cancer molecular sub-
types based on the estrogen receptor (ER) status: ER-positive con-
sists of two main subtypes (Luminal A and B) and ER-negative
consists of three subtypes (Her-2, Basal-like, and Normal-like).
Next, we investigated whether the 71-gene subset extracted
from the 306 intrinsic gene set provides better accuracy in classifi-
cation of subtypes; results are shown in Fig. 7. There are 248
patients in this dataset. SOM produced similar results to those with
the previous dataset indicating similar relative positioning of sub-
types in the SOM panels. For example, Normal-like, Basal-like,
and HER2 subtypes occupy the top left region of the map and
LumA and LumB occupy the bottom left region. This reversal of
space occupied by subsets is not an issue as relative position is what
gives a map its meaning. In this case, different intrinsic genes do
not vary the final SOM result.
In an attempt to find final clusters on SOM, Ward clustering of
SOM neurons provided the Ward index vs. number of clusters rela-
tion for the 93 and 71 gene signatures as shown in Fig. 8. The opti-
mum number of clusters for these two signatures is four and two
clusters, respectively. This indicates that the 93 gene signature has
greater discriminating ability than the 71 gene signature that sees
more similarity between the subtypes leading to only two optimum
clusters. A summary of clustering results for the two gene sets are
given in Table 7 which indicates that 93 gene signature distinguished
the 28 LumA patients but put some Lum B patients with them in
the same cluster. It also separated most (90 %) of Normal-like and
Basal-like into separate clusters. It sees all HER2 patients together in
a cluster but it is shared with 60 % of LumB patients and a small
number of Basal-like and Normal-like patients. The 71-gene subset
in contrast cannot distinguish between LumA and B; identifies all 64
Basal-like patients in one cluster but it shares over 50 % of Normal-
like and HER2 patients. The rest of the Normal-like and HER2
patients are clustered with LumA and B. Clearly, the 93 gene signa-
ture shows greater discriminating ability in this case.
4.2.2 FCM Clustering In this section, we cluster the patients based on the 93 and 71 gene
of Patients signatures using FCM in order to ascertain the efficacy of this clus-
tering method on the selected datasets. Results are shown in
Fig. 7 SOM sample hits (patients) for each “intrinsic” subtype according to the small subset of 71 genes from
the 306 “intrinsic” genes (248 patients in total)
Table 8 where a and b sections are for the 93 gene signature and c
and d parts are for the 71 gene signature. Table 8a, c show original
samples (patient IDs), FCM clusters, and samples in each of these
clusters. In Table 8b, d, FCM clusters are compared with the origi-
nal classification. For the 93-gene set, one FCM cluster has 100 %
members belonging to Normal-like group while other clusters
contain one or more subtypes. Cluster 1 has only LumA patients
a b
5 6
4.5
5
4
3.5
4
3
2.5 3
2
2
1.5
1
1
0.5
0 0
0 10 20 30 40 50 60 70 80 0 50 100 150 200 250
Fig. 8 Ward likelihood index for various number of clusters of map neurons (a) for the 93 subset extracted from
the 546 “intrinsic” genes (b) for the 71-gene subset extracted from the 306 “intrinsic” genes (x-axis indicate
the number of clusters and y-axis-Ward likelihood index). The larger the likelihood index, more likely that
number of clusters
Table 7
Members (patients) in each SOM cluster according to subtype based on: (a) 93-gene subset and (b)
71-gene subset
Patient cluster 1 2 3 4
(a) 93-gene subset from the 546 “intrinsic” genes
HER2 (11 patients) 11
Basal-like (19 patients) 17/19 2/19
Normal (10 patients) 9/10 1/10
Luminal A (28 patients) 28/28
Luminal B (11 patients) 4/11 1/11 6/11
(b) 71-gene subset from the 306 “intrinsic” genes
HER2 (29 patients) 15/29 14/29
Basal-like (64 patients) 64/64
Normal (19 patients) 11/19 8/19
Luminal A (90 patients) 90/90
Luminal B (46 patients) 46/46
but contain only 57 % of them. Similarly, with 71 gene signature,

each cluster consists of two or more subtypes as shown in Table 8d.
It appears that the 93 gene signature based on the original 534
gene list enables more accurate clustering than the 71-gene subset
from the 306 gene signature. To be able to see clearly how patients
are grouped into “intrinsic” subtypes, we projected the FCM
sample cluster results from the 93 gene signature onto our SOM in
Fig. 6 and results are shown in Fig. 9. It shows that Normal-like is
distinct (cluster 5), Basal-like is almost distinct (cluster 4) with an
Table 8
FCM (with 1.2 degree of fuzziness) classification results for two selected datasets: (a) and (b) for 93
genes from the 534 “intrinsic” genes; (c and d) for 71 genes from the 306 “intrinsic” genes
(a) Sample allocation from FCM (based on 93-gene subset from the 534 “intrinsic” genes)
Subtypes Sorlie et al. [8] FCM

LumA (28) Sample : 1–28 Cluster 1 Samples: 1:5,7:8,10:12,14,23:26,28
LumB (11) Sample : 29–39 Cluster 2 Samples: 6,9,13,15:22,27,29,60
HER2 (11) Sample : 40–50 Cluster 3 Samples : 30:36,37,39:50
Basal (19) Sample : 51–69 Cluster 4 Samples: 38,51:59,61:69
Normal (10) Sample : 70–79 Cluster 5 Samples: 70–79
(b) Percentage of patients found in each cluster according to subtype (based on 93 genes from 534
“intrinsic” genes)
Cluster 1 (%) Cluster 2 (%) Cluster 3 (%) Cluster 4 (%) Cluster 5 (%)
LumA (28) 57.15 42.85
LumB (11) 9.09 81.82 9.09
HER2 (11) 100
Basal (19) 5.27 94.73
Normal (10) 100
(c) Sample allocation from FCM (based on 71 genes from 306 “intrinsic” genes)
Subtypes Hu et al. [27] FCM

LumA (90) Sample : Cluster 1 Samples: 1:2,7,10:11,13:17,24,26:27,31,42,45,122,138,
1–46 144:145,147,149,160,162,213:219,221-241,243:249
LumB (46) Sample : Cluster 2 Samples: 3:6,8:9,12,18:23,25,28:30,32:41,43:44,46:47,
46–136 49:50,53,57:58,60:62,70,72:76,78,85,102:105,109,
112:113,126,134,136,242
Normal (29) Sample : Cluster 3 Samples: 48,51:52,54:56,59,63:69,71,77,79:84,86:101,
137–155 106:108,110:111,114:121,123:125,127:133,135,13,
139:140,142:143, 148,150:155
Basal (64) Sample : Cluster 4 Samples: 141,146,156:159,161,163:212,220
156–220
HER2 (19) Sample :
221–249
(continued)
Table 8
(continued)
(d) Percentage of patients found in each cluster according to subtype (based on 71 genes from 306
“intrinsic” genes)
Cluster 1 (%) Cluster 2 (%) Cluster 3 (%) Cluster 4 (%)

LumA (90) 35 65
LumB (46) 1 33 66
Normal (29) 26 64 10
Basal (64) 14 86
HER2 (19) 97 3
Fig. 9 SOM sample hits for each FCM cluster according to the subset of 93 genes from 546 intrinsic genes
odd Lum B patient showing in a distance, all HER2 patients in

cluster 3 with the majority in the middle region of the cluster. The
top region of this cluster contains all LumB patients. LumA is
divided into Cluster 1 and Cluster 2 but overlaying cluster 1 and 2
maps indicate that LumA occupies the top left corner of the map.
What we see here is that Normal-like, Basal-like, and HER2 in that
order occupy the lower diagonal region of the map. Lum A and
Lum B are on the opposite side of the map.
a b c
Basal HER2 Luminal
2006 2006 2006
10 7 17
2 1 1 1
1 3 3
12 4 18 6 1 7 7 5 19
2001 2003 2001 2003 2001 2003

Fig. 10 Venn diagram of original gene clusters: (a) Basal-like, (b) HER2, and (c) Luminal using the small gene
subsets from Sorlie et al. [7, 8] and Hu et al. [27]
We also tested the number of common genes among the

s ubsets from the three studies, Sorlie et al. [7, 8] (2001 and 2003)
and Hu et al. [27] (2006) for three subtypes. As shown in Fig. 10,
there were very few common genes—1, 3, and 3 genes respectively
for Basal-like, HER2, and Luminal. The lack of agreement between
the reported genes continues to pose challenges in identifying the
minimum intrinsic gene subset for distinguishing breast cancer
subtypes.
5 Summary and Conclusions
In this study, an in-depth investigation of the efficacy of three novel

gene sets reported in the literature in distinguishing breast cancer
subtypes was performed using three emergent advanced computa-
tional methods. Selecting computational methods to analyze
experimental data is still a challenging task as there is a myriad of
methods available. Using different methods and distance measures
can produce apparently different results. A good understanding of
the cluster structure of a dataset and what relations are used
whether trends or magnitudes is essential in order to select the
right tools and distance measurement. The three selected tools
(SOM, FLAME, and FCM) were assessed based on the novel gene
signatures proposed by Sorlie et al. [7, 8] and Hu et al. [27] to
investigate the efficacy of 71 and 93 and 71 novel gene subsets,
respectively, proposed by the authors. These small subsets were
extracted by the respective authors from 456, 534, and 306 intrin-
sic genes found differentially expressed in their studies.
First, the two gene sets consisting of 71 and 93 genes charac-
terizing 4 and 5 cancer subtypes, respectively, were analyzed using
the four methods: SOM, ESOM, FLAME, and FCM. SOM and

ESOM are visualization methods for high dimensional data pro-
jected onto a two-dimensional map and they gave an overview of
how genes are organized in terms of distance and density. More
consistent clusters were obtained from FCM and FLAME. FLAME
and FCM with fuzzy characteristics showed that some genes have
expression patterns that characterize two or more subtypes.
We found some similarities and differences in clustering the
selected breast cancer genes with different clustering tools. For the
93-gene subset from the 534 intrinsic genes, SOM gives three
clusters: cluster 1 (LumA), cluster 2 (Basal and Normal-like), and
cluster 3 (LumB, HER2, Basal, and Normal-like); FLAME pro-
duced four clusters: cluster 1 (HER2), cluster 2 (Basal and Normal-
like), cluster 3 (LumA), cluster 4 (LumB); and FCM defined five
clusters: cluster 1 (HER2), cluster 2 (Basal-like), cluster 3 (LumA,
Basal, and Normal-like), cluster 4 (LumA, LumB, and Basal), clus-
ter 5 (LumA and LumB). Thus FLAME clusters genes in almost
the same way as the original results but has difficulty in separating
gene expression patterns from Basal-like and Normal-like subtypes.
FCM has more than one subtype in each cluster except for the one
representing HER2 subtype. SOM also has more than one subtype
in each cluster except for the one representing LumA. For the
71-gene subset from 456 intrinsic genes, FCM had perfect cluster-
ing with one subtype only in each cluster. Next best was FLAME
with all but one cluster representing one subtype each. SOM had
spread each subtype in two clusters. In summary, the 93-gene
subset from 534 “intrinsic” genes gave better subtype classification
when compared with the 71-gene subset from 456 “intrinsic”
genes based on SOM and FLAME but vice versa for FCM which
gave perfect subtype classification for the 71 gene subset. However,
this 71-gene set represented only four subtypes whereas the
93-gene set represented five subtypes. Therefore, the 93-gene set
can provide a further classification of Luminal into Luminal A and
Luminal B.
Before patient clustering based on the Sorlie et al. [8] and Hu
et al. [27] gene sets, we checked and found that there were only a
few overlapping genes from the three different proposed subsets
(Sorlie et al. [7, 8] and Hu et al. [27]) in representing the sub-
types. When using the 93-gene subset from the 546 “intrinsic”
genes, we retrieved better patient classification results from both
SOM and FCM compared with the 71-gene subset from the 306
“intrinsic” genes [26]. In this regard, the main contribution is
using SOM to show the relationship/similarity between subtypes
through visualization. In this subtype classification, we first used
SOM and investigated the spatial organization of original subtypes
based on the 93 and 71-gene subsets. We found that both gene
sets provide consistent results, revealing that each subtype occupies
a distinct location on the map. For example, SOM sample hits
(Figs. 6 and 7) show the opposite positioning of Basal-like and
Luminal A as well Normal-like and Luminal B; HER2 always

appeared in a distinct central position but closer to Normal-like
and Basal-like than either LumA or B. Furthermore, LumA and
LumB together occupy the opposite diagonal area from that occu-
pied by the combined HER2, Basal-like, and Normal-like sub-
types. These and all clustering results indicated the similarity
between the Normal-like and Basal-like subtypes.
We further clustered the SOM using the Ward method and
found that the four optimum clusters produced by the 93-gene set
are more similar to the original subtypes than the two clusters pro-
duced by the 71-gene set from the Hu et al. [27] study. Additionally,
as a validation, we used FCM to cluster original gene expression
profiles with 93 and 71 genes and projected the resulting patient
clusters on to the SOM which showed that the 93-gene subset had
more discriminating ability than the 71-gene set. Thus the pro-
posed 93-gene set demonstrated its strength in differentiating sub-
types to a reasonably high degree of accuracy. There is evidence
that this subset can represent some subtypes such as HER2 well
and do reasonably well for LumA and B, indicating that it has cap-
tured some essential genes. There still is overlap between some
subtypes such as Normal-like and Basal-like.
Given the costly procedures involved in clinical studies, the
proposed 93-gene set can be used for preliminary classification of
breast cancer. The patient based SOM model can be used to map
the gene signature of new patients to locate them on the SOM
with respect to all subtypes to get a comprehensive view of the
classification. This can be followed by a deeper investigation in
light of the observations made in this study regarding overlapping
subtypes. We used gene signatures generated from three compre-
hensive studies conducted in 2001, 2003 and 2006 with one feed-
ing into the other. Our approach and results could be the base for
further refining the gene signatures from later experiments and
those potentially designed to separate as much as possible overlap-
ping clusters as well as to maximally separate all clusters.
Acknowledgements
This research was funded by a Research Grant awarded by Lincoln

University, New Zealand.
References
1. Carey LA (2010) Through a glass darkly: 3. Marchionni L et al (2008) Systematic review:
advances in understanding breast cancer biology, gene expression profiling assays in early-stage
2000–2010. Clin Breast Cancer 10(3):188–195 breast cancer. Ann Intern Med 148(5):
2. Stadler ZK, Come SE (2009) Review of gene- 358–369
expression profiling and its clinical use in breast 4. Desmedt C, Ruiz-Garcia E, André F (2008)
cancer. Crit Rev Oncol Hematol 69(1):1–11 Gene expression predictors in breast cancer:
current status, limitations and perspectives. Eur 19. Weigelt B et al (2012) Challenges translating
J Cancer 44:2714–2720 breast cancer gene signatures into the clinic.
5. Uma RS, Rajkumar T (2007) DNA microarray Nat Rev Clin Oncol 9(1):58–64
and breast cancer – a review. Int J Human 20. Reis-Filho JS, Pusztai L (2011) Gene expres-
Genet 7(1):49–56 sion profiling in breast cancer: classification,
6. Perou C et al (2000) Molecular portraits of prognostication, and prediction. Lancet 378
human breast tumours. Nature 406(6797): (9805):1812–1823
747–752 21. Arpino G et al (2013) Gene expression profil-
7. Sorlie T et al (2001) Gene expression patterns ing in breast cancer: a clinical perspective.
of breast carcinomas distinguish tumor sub- Breast 22(2):109–120
classes with clinical implications. Proc Natl 22. Weigelt B, Peterse JL, van’t Veer LJ (2005)
Acad Sci 98(19):10869–10874 Breast cancer metastasis: markers and models.
8. Sorlie T et al (2003) Repeated observation of Nat Rev Cancer 5(8):591–602
breast tumor subtypes in independent gene 23. Kohonen T (1997) Exploration of very large data-
expression data sets. Proc Natl Acad Sci bases by self-organizing maps. Neural Networks,
100(14):8418–8423 International Conference on 1:PL1–PL6
9. Weigelt B et al (2005) Molecular portraits and 24. Ultsch A, Morchen F (2005) ESOM-Maps:
70-gene prognosis signature are preserved tools for clustering, visualization, and classifica-
throughout the metastatic process of breast tion with Emergent SOM. Technical report 17.
cancer. Cancer Res 65(20):9155–9158 Data Bionics Research Group, University of
10. Henshall S (2005) Review of classification of Marburg, Marburg
human breast cancer using gene expression 25. Fu L, Medico E (2007) FLAME, a novel

profiling as a component of the survival predic- fuzzy clustering method for the analysis of
tor algorithm. Breast Cancer Online 7(12) DNA icroarray data. BMC Bioinformatics
11. Loi S (2008) Review of comparison of gene 8(1):3
expression profiles predicting progression in 26. Dembele D, Kastner P (2003) Fuzzy C-means
breast cancer patients treated with tamoxifen. method for clustering microarray data.
Breast Cancer Online 11(12):null Bioinformatics 19(8):973–980
12. van’t Veer LJ et al (2002) Gene expression pro- 27. Hu Z et al (2006) The molecular portraits of
filing predicts clinical outcome of breast cancer. breast tumors are conserved across microarray
Nature 415(6871):530–536 platforms. BMC Genomics 7(1):96
13. van de Vijver MJ et al (2002) A gene-
28. Sotiriou C et al (2003) Breast cancer classifica-
expression signature as a predictor of survival tion and prognosis based on gene expression
in breast cancer. N Engl J Med 347(25): profiles from a population-based study. Proc
1999–2009 Natl Acad Sci U S A 100(18):10393–10398
14. West M et al (2001) Predicting the clinical sta- 29. Samarasinghe S (2006) Neural networks for
tus of human breast cancer by using gene applied sciences and engineering: from funda-
expression profiles. Proc Natl Acad Sci 98(20): mentals to complex pattern recognition. CRC
11462 Press, Boca Raton, FL
15. Loi S et al (2005) Breast cancer gene expres- 30. Ward JH (1963) Hierarchical grouping to
sion profiling: clinical trial and practice implica- optimize an objective function. J Am Stat
tions. Pharmacogenomics 6(1):49–58 Assoc 58(301):236–244
16. Hess KR et al (2006) Pharmacogenomic pre- 31. Ultsch A, Mörchen F (2005) ESOM-Maps:
dictor of sensitivity to preoperative chemo- tools for clustering, visualization, and classifica-
therapy with paclitaxel and fluorouracil, tion with Emergent SOM. Technical report 46.
doxorubicin, and cyclophosphamide in breast CS Department, University Marburg, Marburg
cancer. J Clin Oncol 24(26):4236–4244 32. Dunn JC (1973) A fuzzy relative of the
17. Naderi A et al (2006) A gene-expression signa- ISODATA process and its use in detecting
ture to predict survival in breast cancer across compact well-separated clusters. Journal of
independent data sets. Oncogene 26(10): Cybernetics 3(3):32–57
1507–1516 33. Bezdek JC (1981) Pattern recognition with
18. Wirapati P et al (2008) Meta-analysis of gene fuzzy objective function algorithms. Kluwer,
expression profiles in breast cancer: toward a Baltimore, MD, p 256
unified understanding of breast cancer subtyp- 34. Dembélé D, Kastner P (2003) Fuzzy C-means
ing and prognosis signatures. Breast Cancer method for clustering microarray data.
Res 10(4):R65 Bioinformatics 19(8):973–980
35. Ozkan I, Turksen IB (2007) Upper and lower survival. Proc Natl Acad Sci U S A 102(10):
values for the level of fuzziness in FCM. Inform 3738–3743
Sci 177(23):5143–5152 40. Chang HY et al (2004) Gene expression signa-
36. Van De Vijver MJ et al (2002) A gene-expression ture of fibroblast serum response predicts
signature as a predictor of survival in breast human cancer progression: similarities between
cancer. New England Journal of Medicine tumors and wounds. PLoS Biol 2(2):E7
347(25):1999–2009 41. Perreard L et al (2006) Classification and risk
37. Huang E et al (2003) Gene expression predic- stratification of invasive breast carcinomas
tors of breast cancer outcomes. Lancet 361 using a real-time quantitative RT-PCR assay.
(9369):1590–1596 Breast Cancer Res 8(2):R23
38. Wang Y et al (2005) Gene-expression profiles 42. Parker JS et al (2009) Supervised risk predictor
to predict distant metastasis of lymph-node- of breast cancer based on intrinsic subtypes.
negative primary breast cancer. Lancet 365 J Clin Oncol 27(8):1160–1167
(9460):671–679 43. Bild AH et al (2009) An integration of comple-
39. Chang HY et al (2005) Robustness, scalability, mentary strategies for gene-expression analysis
and integration of a wound-response gene to reveal novel therapeutic opportunities for
expression signature in predicting breast cancer breast cancer. Breast Cancer Res 11(4):55
Chapter 19
QSAR/QSPR as an Application of Artificial Neural Networks

Narelle Montañez-Godínez, Aracely C. Martínez-Olguín, Omar Deeb,
Ramón Garduño-Juárez, and Guillermo Ramírez-Galicia
Abstract
Quantitative Structure–Activity Relationships (QSARs) and Quantitative Structure–Property Relationships
(QSPRs) are mathematical models used to describe and predict a particular activity/property of com-
pounds. On the other hand, the Artificial Neural Network (ANN) is a tool that emulates the human brain
to solve very complex problems. The exponential need for new compounds in the drug industry requires
alternatives for experimental methods to decrease development time and costs. This is where chemical
computational methods have a great relevance, especially QSAR/QSPR-ANN. This chapter shows the
importance of QSAR/QSPR-ANN and provides examples of its use.
Key words Artificial neural networks, Quantitative structure–activity relationship, Quantitative structure–
property relationship, Principal component analysis
1 Introduction
The first Neural Network (NN) model goes back to 1943 when it
was proposed by Warren McCulloch (1898–1969), a neurophysi-
ologist, and Walter Pitts (1923–1969), a mathematician. The first
attempt to use NN in computers took place in 1956 by a group of
IBM researchers and neuroscientists from McGill University in
Canada. Not only neurosciences have contributed to the develop-
ment of Artificial Neural Networks (ANN). For instance, the
network unit known as a Perceptron was proposed in 1957 by
Frank Rosenblatt (1928–1971), a psychologist, and is still in use
due to its simplicity.
In addition, the development of computer science has fol-
lowed the way of the mathematician John von Neumann. In 1947,
he designed a structure based on sequential processing of data and
instructions. This structure rigorously follows a sequence defined
by data stored in memory. Von Neumann’s architecture is based on
the logic of procedures, normally used for partial solutions in some
problems to link the specific problem with the result. Although the
319
320 Narelle Montañez-Godínez et al.
future of traditional computers seems to be very encouraging,

promising, and complementary to human skills, biological neurons
have attracted the attention of many scientists interested in the
functioning of the brain and its structures, because its operations
offer a possible alternative to computational processors [1].
Nowadays, ANN has reach maturity due to the efforts of many
research groups around the world; they are applied in the solution
of numerous complex real-world problems. ANN and artificial
neural systems in general are multidisciplinary fields, making large
contributions to different areas such as physics, mathematics,
biology, engineering, and others [2]. Their best performance is in
very complex problems that are too difficult for conventional
technology—in other words, problems for which no algorithmic
solution has been found. Some of the areas of application are the
stock market, sales prediction, medical diagnosis, and object recog-
nition, among others.
The best example of biological neural networks is the brain, an
organ made of billions of specialized cells called neurons, intercon-
nected by synapses. The synapses is the zone where two neurons
are connected. Two further important parts of the cell are the
dendrites, which function as the input channel to the cell, and
the axons, which function as the output channel from it. A very
simple representation of a cell is shown in Fig. 1.
The neuron body evaluates the incoming impulses and com-
bines them through a network function that provides the strength of
the neuron. As can be seen an artificial neuron behaves very much
like the biological neuron but at a simplified level. Just as the
biological neuron cannot exists by itself, artificial neurons as inde-
pendent units are not useful for treating information; in order for
them to work they must be grouped into larger structures, the ANN.
Artificial neurons are models that simulate the behavior of
biological neurons. Each artificial neuron is represented as a
Fig. 1 Components of a biological neuron

QSAR/QSPR 321
Fig. 2 Model of an artificial neuron according to McCulloch and Pitts (1943)
processing unit as illustrated in Fig. 2. This illustration shows a

processing unit made of a series of inputs Xi, that are equivalent to
the dendrites that receive the input signals regulated by synaptic
weights represented by wi and a parameter θ that is interpreted as
the minimal threshold that the neuron has to overcome to be
activated. In the neuron body, the output f(y) is processed as a
function of the synaptic weights (w1, w2,…, wn) of each one of the
input data (x1, x2,…, xn) and the threshold θ; mathematically:
f(y) = ∑ i = 1nwi ⋅ xi + θ. Finally, the output value goes through an acti-
vation function that transforms the domain value, possibly infinite,
into a determined value within a specific interval. A biological neu-
ron can be active (excited) or inactive (non-excited); that is, it has
an activation state. In a similar fashion, an artificial neuron also has
different activation states; some artificial neurons can adopt only
two states, in the same way as biological neurons; however others
can take any values within a given interval. In artificial neurons the
activation values are determined by the activation functions, the
most common of which are sigmoidal or logistic and the hyper-
bolic tangent. Both functions can receive input values within the
range [+∞, −∞]; the difference between them is that the first
constrains the output values between [0, 1] and the second
between [−1, +1]. However, there are many activation functions
depending on the application.
The ANN has featured in the scientific field of artificial intelli-
gence because the interaction of information among the processors
depends on the behavior of the whole system. That is, it deals with
understanding from a computerized view of intelligent behavior.
ANN emulates the human brain using a simulation system that is a
computer program; the structures are modulated with high-
capacity parallel computing, known as emulation, or by building
systems similar to biological neural networks, which can learn from
stimuli provided by their environment, whilst artificial neurons

must be programmed by computer based on the fastest micropro-
cessor or Von Neumann’s microprocessor that is capable of reliably
executing complex instructions. ANN can emulate the set of cog-
nitive and intellective skills performed by the human brain. These
ANN are interconnected with a hierarchical organization that
allows performing tasks such as sorting and optimization by the
information of sequential instructions that have been stored in
memory, manipulating data [3, 4].
ANN is a nonparametric, nonlinear modeling technique that
has attracted increasing attention in recent years [5]. Nonlinear
multivariate maps apply a nonlinear transformation of the input
variable space to project inputs onto the designated attribute values
in output space. The power of modeling with layered, feed-forward
artificial neural networks lies in the flexibility of the model defined
by the weights of connections between units within the network.
Together, linear and nonlinear mapping functions may be modeled
by appropriate configuring of the network. Multilayer feed-for-
ward neural networks trained with a back-propagation learning
algorithm have become increasingly accepted approaches.
An important feature of ANN is its dynamic adaptability or
ability to change its behavior in different contexts and with distinct
problems. This is because it relies upon techniques inspired in
learning, generalization or self-organization using elementary pro-
cessing units. This general behavior determines its ability to test
hypotheses, detect statistical patterns and regularities, as well as to
adjust an implicit model which is implemented in the architecture
of the network, which does not depend on the sum of the poten-
tials of neurons [4].
The most common way to implement the models and algo-
rithms adjusted by ANN has been through simulations on conven-
tional computers. Moreover, this can be performed through
architectures that are oriented to the execution of parallel processes
completed on a set of processors connected with some regularity
operating concurrently (neuro-computers), in this type of model,
we can find the classic models Delta + and Mark [4].
Formally, an ANN can be explained with the concept of the
graph, an object composed of a set of nodes (vertices) and connec-
tions (links) that can be led when all connections are assigned a
path and not led when connections are directional. Heavy graphs
are those in which all nodes are connected to each other, while
dispersed graphs have few associations. The graphs can be com-
posed of different types of connections and nodes.
One way of representing graphs is with circles for the nodes,
and lines or arrows for the connections (see Fig. 3). ANN normally
keeps with certain properties and characteristics as the architecture
of ANN and the modes of operation of the network [3].
QSAR/QSPR 323
Fig. 3 A representation of an ANN
ANN is generally formed by a set of elementary processors that

are called artificial neurons; these constitute simple devices of cal-
culation based on information from other neurons or input from
the outside to provide a unique response (output). According to
their location in the network three types of layers can be distin-
guished: Input Layer—neurons that directly receive the informa-
tion coming from an external source; Output Layer—neurons that
present the processed data; Hidden-Layer—neurons that are nei-
ther in the input layer nor the output layer. These neurons are
essentially hidden from view, and their number and organization
can typically be treated as a black box to people who are interfacing
with the system.
There are a number of different parameters that must be
decided upon when designing a neural network. Among these
parameters are: the number of layers, the number of neurons per
layer, the number of training iterations, etc. Some of the more
important parameters in terms of training and network capacity are
the number of hidden neurons, the learning rate, and the momen-
tum parameter.
The input neurons are those that receive information from the
outside or the environment, through sensors or parts of the sys-
tem. Meanwhile, the output neurons are those that send a signal
directly out of the system when the information obtained has been
analyzed or treated is marked by the incoming pulses; hidden neu-
rons receive stimuli and emit output information but only within
the system, so they have no contact with the outside world, they
carry out the basic processing [3, 4].
The basic Perceptron concept is associated with a sensor, which
might provide information on temperature, moisture, liquid levels,
etc.; alternatively, input data may be fed in from a numerical data-
base. The Perceptron is essentially a device that, given the presence
of input data, can generate a single, recognizable output signal.
However, an improvement can be achieved if we add more input
channels to this simple device; it will then be able to differentiate
the different input data and to deliver an outcome through analysis

of that data.
The successful use of neural networks to solve tasks in different
areas is due to the ease with which an ANN acquires information
or knowledge of a problem, as well as its ability to store information
already acquired for later use, and, of course, its effectiveness in
solving problems from sub-solutions. However, difficulties arise
when we try to understand how a solution has been reached,
because an ANN provides solutions to problems or outputs,
without providing at the same time an explanation as to how that
output was arrived at [6].
2 Methodology
Extraction is an important technique in ANN, even though this

has a cost in resources and effort.
In artificial intelligence, the term of explanation refers to an
explicit structure which can be used internally within, for example,
an ANN, to elucidate and understand information and then explain
the results obtained to the user. The information can be external-
ized with object hierarchy, semantic networks, frames, and so on.
This explanation also includes the steps of the reasoning process or
intermediate steps of the process as the trace of activity rules or test
structure. This can assist the user in checking the internal logic of
the system, a particularly helpful step if the explanations are very
coarse or limited.
In many systems, there are deficiencies in rule extraction
because it is very difficult to generate the explanatory structure of
the process, even though the helpful capacity and quality of expla-
nation are important in use of a trained ANN. The quality of expla-
nation refers to how direct the task of extraction from ANN is. For
example, Rule of trails may be used to capture an explanation, but
it has been shown that this approach is often too rigid or inflexible,
since these rules may contain references about internal calculations
and repetitions [6].
2.1 Software In most examples of the use of ANN in QSAR, the methodology
uses quantum level calculations to optimize the molecular geome-
tries data set, using semi-empirical methods or Density Functional
Theory methods to generate descriptors and the subsequent
selection of them to generate a new source of descriptors. The
principal software tools used are Gaussian [7], MOPAC [8], and
Hyperchem [9]. Numerical analysis can be realized with MATLAB
[10] for linear regressions, an ANN for non-(multi)linear regres-
sion and SPSS [11] for multilinear regressions. Furthermore,
Dragon [12] software is used in the calculation of descriptors.
QSAR/QSPR 325
3 QSAR/QSPR
For both financial and social reasons, the drug industry would like
to develop new drugs in much shorter times than is possible using
a purely experimental (lab-based) approach. A large number of
novel compounds are synthesized each year. The number of
molecules that can be obtained by organic synthesis is so high that
the probability of choosing randomly a molecule with the desired
biological activity is practically nil. In this regard, medicinal chem-
istry techniques that allow discovery and optimization of new
templates are used. However, a large fraction of these compounds
are not tested for physicochemical properties or biological activi-
ties, which still remain unknown owing to unavailability of the
compounds or possible risks of toxicity. A method able to predict,
within a realistic error boundary, the biological activities of untested
compounds is necessary to evaluate these molecular features in a
rapid and inexpensive approach.
Among these methods, computer aided design has found its
way to the rational design of new compounds; the most important
practice is known as Quantitative Structure–Activity Relationship
(QSAR). In recent years, numerous quantitative structure–activ-
ity/property relationship QSAR/QSPR models have been intro-
duced for calculating the biological activities from a variety of
numerical descriptors of chemical structures. These relationships
develop correlations between the descriptors of individual com-
pounds and their biological activity/chemical properties. This
approach has proved especially productive when coupled with solid
optimizing tools that help to narrow down the field of potential
molecules to just the most promising candidates, namely, Genetic
Algorithms, Ant Colony Optimization, Support Vector Machines,
and Artificial Neural Networks.
An inevitable difficulty, when dealing with molecular descrip-
tors, is that of collinearity, which frequently exists between inde-
pendent variables. In the worst case, this creates a rigorous problem
in certain types of mathematical treatment such as matrix inversion
[13], though collinearity of molecular descriptors is not normally
so complete that matrix inversion will fail. A better predictive
model can be obtained by orthogonalization of the variables by
means of principal component analysis (PCA) and the subsequent
method is called principal component regression (PCR) [14].
In order to decrease the dimensionality of the independent vari-
able space, a restricted number of principal components (PCs) are
used. For this reason, the choice of significant and informative PCs
is an important task in PCA–based calibration methods [15, 16].
Diverse methods have been proposed to select the important
PCs for calibration purposes. The simplest and most frequent one
is a top–down variable selection where the factors are ranked in the
order of decreasing eigenvalue. The factor with the highest eigen-

value is considered as the most important one; further factors are
then progressively introduced into the calibration model until no
significant improvement of the calibration model is obtained. In
another method, correlation ranking, the factors are ranked by their
correlation coefficient with the property to be correlated (i.e., bio-
logical activity) and selected by the procedure of eigenvalue ranking.
QSAR is a mathematical model that attempts to relate the
structure-derived features of a compound to its biological or physi-
cochemical activity and it is used to understand drug action as well
as to design new compounds. Elaboration of QSAR models starts
with the collection of a representative set of molecules with the
desired biological activity, and a control group that does not pos-
sess the activity under study. This is followed by the selection of
molecular descriptors obtained by a mathematical logic that trans-
forms the chemical and/or structural information about a mole-
cule into a set of numbers. In favorable cases, the molecular
descriptors can then be related to the desired biological activity. In
order to validate the descriptors, the full data set is divided into a
training set, and a learning test set. In the learning process, there
are many ways of building a model, among which we can mention
Multiple Linear Regression (MLR), logistic regression, or machine
learning methods. The optimal model is obtained from the model-
ing parameters and features. Finally, this model is validated to
ensure that it will be useful (see Fig. 4) [17].
A QSAR model correlates, within congeneric molecules, many
biological activities such as inhibition constants, affinities of ligands
Data
Features
Training set Testing set
Learning
Model Validation
Fig. 4 General workflow of developing a QSAR model [17]

QSAR/QSPR 327
to their binding sites and rate constants, with either structural fea-
tures or with atomic group or molecular properties such as solubil-
ity, lipophilicity, electronic effects, ionization, and stereochemistry.
The structural features were proposed by Free and Wilson [18]
and the use of molecular properties was proposed by Hansch and
Fujita [19], both published independently in 1964. QSAR has
been important in the integration of ADMET (absorption, distri-
bution, metabolism, excretion, and toxicity) profiling and predic-
tion, which is mostly dependent on a type of molecular descriptors
as described by the Lipinski Rule of 5. This procedure is applied to
improve our understanding of structure–activity links, to identify
hazardous compounds at low cost, and to reduce the number of
animals used for experimental toxicity testing. In the field of QSAR
and in the prediction of animal toxicity, various artificial intelli-
gence techniques have been proposed and developed such as ANN
and statistical understanding of knowledge networks [20–22].
There are different tools that have been used to predict the
activity of molecules that can present some biological activity; MLR,
nonlinear regression using ANN and molecular simulations from
molecular recognition (Docking) are also used in QSAR [23].
The calculation of chemical descriptors is an important tool for
QSAR applications, because it makes possible prediction of toxic
properties, harmful characteristics, and pharmacological biological
properties that a group of molecules could present. Many chemical
descriptors can be measured experimentally but this is cost and
time expensive, and therefore development of QSAR tools is useful
to replace or augment measurement.
Most QSAR/QSPR practice uses quantum-mechanical descriptors.
QSAR analysis uses a variety of descriptors including:
● Topological,
● Quantum chemical,
● Galvez topological charge index,
● 2D-autocorrelation,
● Radial Distribution Functions (RDF),
● 3D-MoRSE (Molecule Representation of Structures based
on Electron diffraction),
● Weighted Holistic Invariant Molecular (WHIM),
● GEometry Topology and Atom-Weight AssemblY
(GETAWAY), and so on [17, 24].
3.1 QSAR-ANN In the literature, we can find work that reports applications in
Application which only QSAR has been used, e.g., a QSAR model that describes
the antispasmodic activity of molecules isolated from Mexican
Medicinal Flora and some synthetic based on stilbenoid bioiso-
steres [25]. Equally, there are reports of applications that use only
ANN, such as the unsupervised mapping of samples [26], the

prediction of the maximum power of a low concentration photo-
voltaic module [27], or drug analysis [28].
The importance of ANN-QSAR methods is their flexibility to
cover a range of applications. In addition they have the advantage
of providing results with higher speed and lower costs than experi-
ments. Furthermore, the ability of ANN to discover more complex
relationships has allowed this method to find a wide application in
QSAR studies [29, 30]. A principal component–artificial neural
network (PC-ANN) method, which combines the PCA with ANN,
is another PCA-based calibration method for nonlinear modeling
between the PCs and biological activities [13]. Many of these
QSAR studies have been performed in our group.
As an initial example, it is possible to predict toxicity using
QSAR algorithms. In this example, 278 substituted benzenes were
selected and the binding affinities to human serum albumin (HSA)
were modeled for 94 these compounds. HSA is the most abundant
protein in plasma and has a high capacity to bind to several drugs.
We have compared SR-PC-ANN (stepwise regression–principal
components–artificial neural networks) and CR-PC-ANN (corre-
lation ranking–principal components–artificial neural networks)
procedures, both combined with two PCA approaches, the indi-
vidual PCA approach, PCA(I), and the combined PCA approach,
PCA(C). More accurate results were obtained with CR-PC-ANN
and the PCA(I) approaches [31].
As a second example, it is possible to describe and predict the
antinociceptive activity of morphinans derivatives (see Fig. 5) using
QSAR analysis. MRL and Leave-One-Out Cross-Validation
(LOO-CV) were applied to the best QSAR models. Furthermore,
in this instance 31 morphinan derivatives reported by the US Drug
Fig. 5 Different changes of morphine structure localized on the 31 morphinan

studies [29]
QSAR/QSPR 329
a 4
MLR-M86A b ANN-M66B
Training (86 molecules) 4 Training (66 molecules)
Predicted (20 molecules) Predicted (40 molecules)
3 3 NIH11179
NIH11167
NIH11100 NIH11181
NIH11179
2 NIH10945 NIH11188 2 NIH11168
NIH11154 NIH11178
NIH11178
NIH11154
exp
exp
1
log K m
NIH11168 1
log K m
NIH11068 NIH11072
0 0
–1 –1
NIH 11081
–2 –2
–2 –1 0 1 2 3 4 –2 –1 0 1 2 3 4
exp exp
log K m log K m
Fig. 6 Examples of data set calculated by Multiple Linear Regression (solid line), ideal line (dashed line), and
residual line limits (dotted line, cutoff = |1.0|). These were the best ANN calculated results [32]
Evaluation Committee were analyzed. The models calculated

achieved good descriptive and predictive power. This was improved
by using ANN; however it was unable to predict an external valida-
tion set of morphinan derivatives [29].
Third, it is feasible to study biological activity of 106 morphinan
derivatives (reported by the US Drug Evaluation Committee) to
describe μ-receptor-binding affinity. Twenty one descriptors were
selected to perform the MLR. ANN was then used to improve the
results (see Fig. 6). The context of this work is that opiates often
used for the control of pain can also have toxic side effects. There
are different types of opioid receptors, including μ-receptor that
shows the highest affinity for morphine, so the affinity in this site is
an interesting topic of study [29, 32].
Fourth, a QSAR analysis has been conducted on analgesic
activity (log IC) for 95 heterogeneous analgesic compounds using
the PC-ANN modeling method, in which we applied the eigen-
value ranking factor selection procedure. This study demonstrates
that the PC-ANN can be used to generate improved general mod-
els for heterogeneous data sets without splitting them into catego-
ries. The PC-ANN gives better regression models with good
prediction ability using a relatively low number of principal com-
ponents. A 0.834 correlation coefficient was obtained using
PC-ANN with six extracted principal components [33].
Fifth, a QSAR analysis has been performed on three types of
Carbonic Anhydrase (CA) I, II and Bovine IV isozyme inhibitory
activities for 53 sulfonamides using multiple linear regression
(MLR), principal component artificial neural network (PC-ANN),
and correlation ranking–principal component regression (CR-PCR)
analyses. It was observed that the interaction between the
ligand and receptor varies from one type of CA isozyme to another.
The results obtained provide very good regression models with

good prediction ability. Correlation ranking-principal component
regression analysis provides models with better prediction capabil-
ity for the three types of the CA isozyme inhibitory activity, while
PC-ANN analysis provides models with better prediction capabil-
ity for the hCAII isozyme activity. Generally, the models obtained
for modeling the hCAII isozyme inhibitory activity are superior
over those obtained for modeling the hCAI and bCAIV isozyme
inhibitory activities. The QSAR model obtained for CAI indicated
that the molecules with higher hydration energies were found
to bind strongly to the receptor. In addition, it was found that
higher molecular connectivity tends to block the drug binding to
the receptor.
By contrast, for the binding of sulfonamides to CAII isozyme,
it was found that the presence of such rings in the sulfonamides
blocks their binding to the receptor. For the CAIV isozyme, the
obtained QSAR model illustrates the extremely significant role of
softness, so that the more polarizable is the ligand, the stronger it
binds to the receptor. It was found that higher molecular connec-
tivity tends to block the drug binding to the receptor for the three
types of CA isozyme inhibitory activities. The results obtained
show that linear and nonlinear regression analyses are useful tools
to distinguish between the inhibitory activities of sulfonamides
toward different CA isozyme types I, II, and IV [34].
Sixth, a QSAR study was performed to understand the inhibi-
tory activity of a set of 192 vascular endothelial growth factor
receptor-2 (VEGFR-2) compounds. QSAR models were devel-
oped using MLR and PLS as linear methods, while PC-ANN mod-
eling method with application of eigenvalue ranking factor selection
procedure was used as a nonlinear method. The results obtained
offer good regression models having good prediction ability. The
results obtained by MLR and PLS are close, and better than those
obtained by principal component- artificial neural network.
The best model obtained had a correlation coefficient of 0.87. The
strength and the predictive performance of the proposed models
were verified using both internal (cross-validation and Y-scrambling)
and external statistical validations [35].
Seventh, Counter propagation neural network (CPNN) is an
attractive classification tool (active and non-active) in QSAR stud-
ies. A major obstacle in classification by CPNN is finding the best
subset of variables. In this study, the performance of some different
feature selection algorithms including F score-based ranking,
eigenvalue ranking of PCs obtained from the data set, Non-Error-
Rate (NER) ranking of both descriptors and PCs, and 3-way
handling of data, Parallel Factor Analysis (PARAFAC), were evalu-
ated in order to find the best classification model. The methods
were applied for modeling protein-tyrosine kinase inhibition of
105 flavonoid derivatives using substituent electronic descriptors
QSAR/QSPR 331
(SED) as a novel source of electronic descriptors. The results

showed that the best performance was achieved by F-score rank-
ing, while the NER ranking of principal components (PCs) showed
very fluctuating results; the worst performance belonged to
PARAFAC-CPNN. Furthermore, comparison of results of these
nonlinear algorithms with a linear discriminate analysis method
revealed better predictions by the former [36].
3.2 QSPR-ANN Quantitative structure–property relationship studies have also been

Application performed using PC-ANN in our group.
First, a QSAR analysis has been conducted on bioconcentration
factor (BCF) for 227 different non-ionic organic compounds. The
terms bioaccumulation and bioconcentration refer to the uptake
and buildup of chemicals that can occur in living organisms.
Experimental measurement of bioconcentration is time-consuming
and expensive, and is not feasible for a large number of chemicals
of potential regulatory concern. A highly effective tool depending
on a quantitative structure–property relationship (QSPR) can be
used to describe the tendency of chemical concentration organisms
represented by the important ecotoxicological parameter, the loga-
rithm of Bio Concentration Factor (log BCF) with molecular
descriptors for a large set of non-ionic organic compounds. QSPR
models were developed using multiple linear regression, partial
least squares, and neural networks analyses. Linear and nonlinear
QSPR models to predict log BCF of the compounds were devel-
oped for relevant descriptors. The results obtained offer good
regression models with good prediction ability. The descriptors
used in these models depend on the volume, connectivity, molar
refractivity, surface tension, and the presence of atoms accepting
H-bonds [37].
Second, a quantitative-structural property relationship analysis
has been performed on the logarithm of gas/particle partitioning
coefficient (logKp) for 70 different semi-volatile organic com-
pounds by using the PC-ANN modeling method, with application
of an eigenvalue ranking factor selection procedure. The PC-ANN
gives good regression models with good prediction ability using a
relatively low number of PCs. The optimal models obtained by
PC-ANN and partial least square (PLS) analyses are in close prox-
imity from the statistical point of view. The results obtained offer
excellent regression models that hold good prediction ability com-
pared to other studies on the same data set of compounds. A coef-
ficient of determination around 0.97 was obtained using PC-ANN
and PLS analysis [38].
Third, a quantitative structure–property relationship analysis
has been performed on the logarithm of solubility (log 1/S) in
water for 219 different pesticide compounds by using the PLS method
as well as a PC-ANN method, with application of the eigenvalue
ranking factor selection procedure. The PLS and PC-ANN give
good regression models with good prediction ability using a relatively

low number of PCs. The optimal models obtained by PC-ANN are
better than those obtained by PLS analyses from the statistical
point of view. The results obtained offer excellent regression models
that hold good prediction ability. The optimal PLS model has coef-
ficients of determination of 0.7998 and 0.7944 for the training
and test sets, respectively, while the PC-ANN model has coeffi-
cients of determination of 0.8435 and 0.8193 for training and test
sets, respectively. The descriptors used in these models are consis-
tent with the experimental factors that are presumed to affect the
solubility of pesticide compounds in water [39].
4 Conclusions and Perspectives
Currently, QSAR analysis has not reached a level at which the pre-
dictive power is as good as the descriptive power, but when ANN
is included in the models, the predictive power is improved, com-
pared to a model that uses only MLR. When using better comput-
ers and models, predictions will be closer to reality.
Acknowledgement
AC Martínez-Olguín wishes to thank CONACyT for a graduate

scholarship. The English was kindly reviewed by Miss Désirée
Argott.
References
1. de la Fuente Aparicio Ma J, Calonge Cano T model for HEPT derivatives. J Chem Inf
(1999) Aplicaciones de las redes de neuronas Comput Sci 43:1200–1207
en supervisión, diagnosis y control de pro- 6. Andrews R, Diederich J, Tickle AB (1995)
cesos. Universidad Simón Bolívar, Caracas- Survey and critique of techniques for extract-
Venezuela, p 11 ing rules from trained artificial neuronal
2. Brégains JC (2007) Análisis síntesis y control networks. Knowl Base Syst 8:373–389
de diagramas de radiación generados por dis- 7. Frisch MJ, Trucks GW, Schlegel HB et al
tribuciones continuas y discretas utilizando (2009) Gaussian 09. Gaussian, Inc.,
algoritmos estadísticos y redes neuronales arti- Wallingford CT
ficiales aplicaciones. Universidad de Santiago 8. Stewart JJP (2009) MOPAC 2009. Fujitsu
de Compostela, Galicia-Spain, p 104 Limited, Tokyo, Japan
3. Flores López R, Fernández Fernández JM 9. HyperChem(TM) 2003 Professional 7.51,
(2008) Las redes neuronales artificiales Hypercube, Inc., 1115 NW 4th Street,
Fundamentos teóricos y aplicaciones prácticas. Gainesville, Florida 32601, USA
Netbiblo, S.L., Spain, pp 12–14, 17, 21
10. MATLAB Version 7.2.0.232 (R2006a). The
4. Pino Diez R, Gómez Gómez A, de Abajo MathWorks, Inc. (2006) Natick, MA
Martínez N (2001) Introducción a la inteligen-
cia artificial: Sistemas expertos, redes neuronales 11. SPSS Version 13 for Windows (2004)
artificiales y computación evolutiva. Universidad Chicago, IL
de Oviedo, Asturias-Spain, pp 27, 28 12. Todeschini R, Consonni V, Pavan M (2002)
5. Douali L, Villemin D, Cherqaoui D (2003) Dragon Software version 2.1-2002 Pisani 13,
Neural networks: accurate nonlinear QSAR Milano. Dragon Software and references therein
QSAR/QSPR 333
13. Vendrame R, Braga RS, Takahata Y et al power of a low concentration photovoltaic

(1999) Structure–activity relationship studies module for building integration. Sol Energ
of carcinogenic activity of polycyclic aromatic 100:148–158
hydrocarbons using calculated molecular 28. Agatonovic-Kustrin S, Beresford R (2000) Basic
descriptors with principal component analysis concepts of artificial neural network (ANN)
and neural network methods. J Chem Inf modeling and its application in pharmaceutical
Comput Sci 39:1094–1104 research. J Pharm Biomed Anal 22:717–727
14. Jolliffe IT (1986) Principal component analy- 29. Ramírez-Galicia G, Garduño-Juárez R,
sis. Springer, New York, NY Hemmateenejad B et al (2007) QSAR Study
15. Xie YL, Kalivas JH (1997) Evaluation of prin- on the antinociceptive activity of some mor-
cipal component selection methods to form a phinans. Chem Biol Drug Des 70:53–64
global prediction model by principal compo- 30. Schbeider G, Wrede P (1998) Artificial neural
nent regression. Anal Chim Acta 348:19–27 networks for computer-based molecular dos-
16. Depczynski U, Frost VJ, Molt K (2000) ing. Prog Biophysics Mol Biol 70:175–222
Genetic algorithm applied to the selection of 31. Deeb O (2010) Correlation ranking and step-
factors in principal component regression. wise regression procedures in principal compo-
Anal Chim Acta 420:217–227 nents artificial neural networks modeling with
17. Dehmer M, Varmuza K, Bonchev D (2012) application to predict toxic activity and human
Statistical modelling of molecular descriptors serum albumin binding affinity. Chemometr
in QSAR/QSPR. Wiley, Germany, p 2 Intell Lab Syst 104:181–194
18. Free SM, Wilson JA (1964) A mathematical 32. Ramírez-Galicia G, Garduño-Juárez R, Deeb
contribution to structure-activity studies. J O et al (2008) MLR–ANN and RTO Approach
Med Chem 7:395–399 to μ-opioid receptor-binding affinity. Pooling
19. Hansch C, Fujita T (1964) p-σ-π Analysis. A data from different sources. Chem Biol Drug
method for the correlation of biological activ- Des 71:260–270
ity and chemical structure. J Am Chem Soc 33. Deeb O, Drabh M (2010) Exploring QSARs
86:1616–1626 of some analgesic compounds by
20. Hansch C, Leo A, Hoekman D (1995) PC-ANN. Chem Biol Drug Des 76:255–262
Exploring QSAR: fundamentals and applica- 34. Deeb O, Goodarzi M, Khadikar PV (2012)
tions in chemistry and biology. American Quantum chemical QSAR models to distin-
Chemical Society, Washington, DC guish between inhibitory activities of sulfon-
21. Hansch C, Leo A, Hoekman D (1995) amides against human carbonic anhydrase I
Exploring QSAR: volume 2: hydrophobic, and II and bovine IV isozymes. Chem Biol
electronic, and steric constants. American Drug Des 79:514–522
Chemical Society, Washington, DC 35. Deeb O, Jawabreh S, Goodarzi M (2013)
22. Helma C, Kramer S (2003) A survey of the Exploring QSAR’s of vascular endothelial
predictive toxicology challenge 2000–2001. growth factor receptor-2 (VEGFR-2) tyrosine
Bioinformatics 19:1179–1182 kinase inhibitors by MLR, PLS and
23. Waterbeemd H, Gifford E (2003) ADMET in PC-ANN. Curr Pharm Des 19:2237–2244
silico modeling: towards prediction paradise? 36. Hemmateenejad B, Mehdipou A, Deeb O et al
Nat Rev Drug Discov 2:192–204 (2011) Toward an optimal approach for vari-
24. Karelson M (2000) Molecular descriptors in able selection in counter-propagation neural
QSAR/QSPR. Wiley-Interscience, New York, networks: modeling protein–tyrosine kinase
USA inhibitory of flavanoids using substituent elec-
25. Ramírez-Galicia G, Garduño-Juárez R, tronic descriptors. Mol Inform 30:939–949
Hemmateenejad B et al (2007) QSAR study on 37. Deeb O, Khadikar PV, Goodarzi M (2010)
the relaxant agents from some Mexican medici- QSPR modeling of bioconcentration factors of
nal plants and synthetic related organic com- nonionic organic compounds. Environ Health
pounds. Chem Biol Drug Des 70:143–153 Insights 4:33–47
26. Marini F, Bucci R, Magrì AL et al (2008) 38. Deeb O, Khadikar PV, Goodarzi M (2011)
Artificial neural networks in chemometrics: Study of gas/particle partitioning coefficients
history, examples and perspectives. Microchem of semivolatile organic compounds via QSPR
J 8:178–185 methods. J Iran Chem Soc 8:176–192
27. Fernández EF, Almonacid F, Sarmah N et al 39. Deeb O, Goodarzi M (2010) Predicting the
(2014) A model based on artificial neuronal solubility of pesticides compounds in water via
network for the prediction of the maximum QSPR methods. Mol Phys 108:181–192
INDEX
A Bioactive peptide ......................................................101–116

Bioactivity................................................... 72, 125, 127–129
Accelrys ............................................................................126 Biochemical process..........................................................1, 2
Activation function .............................54, 116, 133, 196, 214, Biochemistry ................................................................13, 15
215, 249, 251, 271, 272, 276, 319 Bioconcentration ........................................................ 81, 329
Actuator............................................................................182 Bioinformatics .....................................48, 124, 125, 274, 275
Adaboost .................................................................. 232, 235 Biological activity ...........................46, 47, 51, 111, 149–162,
Adaptive learning ..................................................... 150, 270 323–325, 327
Adaptor .................................................................. 2–4, 6–14 Biological mechanism.......................................................261
ADMET ..................................... 46, 48, 55, 56, 83, 155, 325 Biology ........................................................... 1, 15, 125, 286
Amino acid ..............3–6, 8–10, 12, 13, 15, 21–23, 26, 28, 89, Biometric authentication ..........................................205–224
101–104, 106–108, 121, 172 Biometric feature set.........................................................207
AMPlist.................................................... 104, 106, 107, 111 Biometric identity.............................................................207
Analgesic activity ..............................................................327 Biopsy ....................................................................... 200, 203
Ant-Colony Optimization................................................323 Bitmap image ...................................................................209
Antibacterial activity.....................................................55–57 Black-box ................................................................... 69, 136
Antibacterial agent .......................................................45–61 Blood type ................................................................ 278, 282
Antibiotic ................................................45, 46, 80, 103, 106 Boiling-point ...........................................70–72, 83, 120, 121
Antibody.............................................................................11 Boltzmann learning rule ........................................... 273, 274
Antimicrobial ......................................49, 103–107, 113, 114 Boosting ...........................................................................232
Architecture .........................9, 53, 58, 95, 122, 133, 134, 164, Bootstrap ............................................................................40
167, 170, 183, 196, 197, 202, 203, 217, 218, 236, Bottom-up method...........................................................200
237, 239, 249, 252, 255, 257, 261, 269–282, Bray-Curtis similarity ...................................................36, 40
317, 320 Breast cancer............................................. 195, 203, 285–315
Artificial pancreas .............................................................246
Artificial synapse ......................................................261–267 C
Authentication..........................................................205–224
Camera ...................................... 206, 208, 229, 231, 234, 253
Autocorrelation....................................82, 108, 184, 210, 325
Cannabinoid (CB) .................................... 150, 151, 154–161
Autoencoder network .......................................................237
Carbohydrate ....................................................................247
AutoWeka ................................................................119–145
Carcinogenicity ........................................ 70, 76, 81–83, 123
CARET.................................................................... 111, 112
B
Cargo .................................................................... 2–8, 10–14
Babel......................................................................... 126, 161 Cargo localisation ...............................................................10
Backbone .......................................................... 18–21, 24, 26 Catalytic function .................................................................1
Backpropagation ............................................... 200, 252, 253 CD-Hit ............................................................................106
Bacteria................................................33, 45, 46, 55, 61, 105 Cell .................................2, 3, 6, 13–15, 46, 48, 161, 270, 318
Bacterial community.....................................................33–42 Cepstral ............................................................................210
Bayesian network ................................................................37 CGM sensor ......................................246–249, 253, 254, 257
Bayesian Network Inference tool (BANJO) ..... 37, 38, 41, 42 Character recognition .......................................................274
Bayesian regularization .....................................................253 Charge prediction .........................................................89–99
Bayesian smoothing ..........................................................250 Charge transfer ...................................................................91
Benzene ...................................................................... 80, 120 Chemical shift ..............................................................17–31
Binding activity ........................................................154–160 Chemogenomics ...............................................................150
Binning....................................................................... 35, 128 Chemotherapy .......................................................... 274, 281
DOI 10.1007/978-1-4939-2239-0, © Springer Science+Business Media New York 2015
335
ARTIFICIAL NEURAL NETWORKS, METHODS IN MOLECULAR BIOLOGY
336 Index
Chimera........................................................................11, 13 Protein Data Bank ................................................ 18, 165

Classification ...................... 21, 24, 26–28, 30, 46, 51–55, 59, PubChem ...................................................................126
60, 67, 77, 79, 82–84, 96, 104, 105, 107, 111–115, SGD genome................................................................12
130, 134, 196, 199, 213, 215, 232, 234, 235, 239, WOMBAT .................................................................126
274–276, 285–315, 328 Data mining ................................................. 54, 55, 119–145
Classifier Degradation............................................................ 76, 79, 81
strong .................................................................. 232, 234 Devtools ...........................................................................104
weak....................................................................232–234 Diabetes.................................................... 150, 246–248, 274
Cluster ......................................... 4, 9, 24, 199, 212, 285–315 Dimensionality .......................... 196, 199, 232, 235, 239, 323
Clustering ................................................20, 23, 55, 285–315 Discretization ...............................................................41, 42
Coil..................................................21, 23, 24, 101, 102, 275 Disease ....... 3, 4, 14, 45, 46, 61, 246–248, 275, 276, 281, 286
Collision-induced dissociation (CID) .............. 89–95, 97–99 DNA .......................................................................... 34, 276
Combinatorial chemistry ....................................................46 DNA-microarray ..............................................................287
Combinatorial library ...........................................................9 Domain ...........................3, 6, 7, 12, 15, 67–69, 79, 121, 124,
CoMFA ...................................................................... 49, 153 209, 210, 218, 257, 300, 319
Compartment ................................................2–4, 6, 8, 10, 15 DRAGON ......................................................... 52, 127, 322
Competitive learning rule ......................................... 273, 274 Drug
Computational neuroscience ............................................227 antibacterial ................................................ 46, 48, 54, 61
Computer-aided diagnosis (CAD) ...........................195–203 development time .......................................................103
COMSIA ........................................................... 49, 153, 154 discovery .......46, 48, 53, 61, 119, 149, 150, 157, 274, 275
Conformation ..................................7, 18, 19, 21, 49–51, 156 resistance ............................................................ 275, 281
Confusion matrix...................................................... 113, 114 Drug-resistant bacteria ............................................... 45, 281
Controller .................................................................180–182 Durbin-Watson test..........................................................278
CORINA .........................................................................126
Correlation coefficient .................. 40, 42, 59, 60, 90, 94, 114, E
153, 210, 304, 324, 327, 328 Early stopping .......................................... 152, 202, 252, 253
Corrplot............................................................................104 Ecosystem...........................................................................34
COSMOS ........................................................................126 Eigenvalue ................................................ 153, 324, 327–329
Cost function ............................................ 111, 197, 200, 213 Eigenvalue ranking ........................................... 324, 327–329
Counter-propagation neural network (CPNN) ........ 328, 329 Electronic synapse ............................................................261
Crisp clustering ................................................................285 Embedded system............................................. 205, 207, 217
Cross-validation ........... 68, 98, 111–113, 116, 152–155, 168, Embryo.................................................................................1
251–253, 278, 326, 328 Emergent properties ...........................................................14
Curation ......................................................... 46, 55–57, 145 Emergent Self-organising Map (ESOM) ............... 288, 290,
Curcumin .......................... 124, 128, 129, 131, 132, 136–138 294–305, 313
Cybenko theorem .............................................................276 Environmental interaction network (EIN) ....... 34, 35, 37–42
CyberHand.......................................................................180 Error function...................................................................134
Etching ..................................................... 262, 263, 265, 266
D Euclidean distance .................................................... 213, 291
Database Eukaryotic cell ..................................................................1, 2
AAindex .....................................................................108 Eureqa .............................................................. 37, 38, 40, 42
ACE ................................................................... 150, 151 European Union (EU) ................................................ 65, 117
AchE .................................................................. 150, 151 Evolutionary algorithm ......................................................37
APD2 .........................................................................106 Excel .............................................................................47, 48
BindingDB .................................................................125 Extended-Connectivity Fingerprint (ECFP6) ........ 152, 158,
BZR.................................................... 150, 151, 153, 154 160, 161
CHEMBL...................................47, 48, 55–57, 126, 156
F
COX2 ......................................................... 150, 151, 153
DHFR ................................................ 150, 151, 153, 154 False acceptance rate (FAR) .............................................206
DrugBank ............................................................. 48, 126 False rejection rate (FRR).................................................206
flybase ...........................................................................12 FANN ......................................................................149–162
LipidBlast .....................................................................90 FANN-QSAR..........................................................149–162
NCBI protein .................................................................5 FCM. See Fuzzy C-means (FCM)
NCI ............................................................ 150, 157–160 Feature extraction ..............199–200, 207, 209–212, 218–219
Index
337
Feedback ........................................3, 109, 110, 179–181, 272 Hidden-Markov model ....................................................281
Feedforward...................................................... 183, 196, 239 Hidden neuron .................. 115, 152–155, 249, 251, 252, 321
Feedforward network........................................................183 Hierarchical clustering ............................. 286, 288–290, 305
Filter .................................... 58, 105, 162, 170–172, 232, 239 High-throughput screening ................................................46
Fingerprint ............................. 51, 77, 89, 149–162, 206–209, Histology ..........................................................................203
211–215, 217–224 Horizon ...................................................... 61, 182, 185, 248
Fingers...................................................... 180–182, 187, 190 Human serum albumin (HSA).........................................326
5D-QSAR ..........................................................................49 Hydrogen bonding .............................................................18
FLAME. See Fuzzy Clustering by Local Approximation of HyperChem ............................................................. 126, 322
Memberships (FLAME) Hyperglycemia.......................................... 246–248, 256, 257
Flash-point ...................................................................70, 73 Hyperglycemic event ........................................................256
Food and Drug Administration (FDA) ..............................48 Hypoglycemia........................................................... 246, 257
4D-QSAR .................................................................. 49, 122 Hypoglycemic event .........................................................256
Fragmentation pattern ........................................................89
Free-Wilson........................................................................49 I
Fuzzy clustering........................................................285–315 Image analysis........................................................... 274, 275
Fuzzy Clustering by Local Approximation of Memberships Infomax ....................................................................239–241
(FLAME) ................289, 290, 295–306, 312–314 Information code ............................................................7, 14
Fuzzy C-means (FCM) ............................ 289, 290, 295–314 Inhibitory effect ....................................................................8
Fuzzy logic.................................206, 207, 215–217, 223, 224 Input layer ...........................9, 54, 95, 96, 151, 152, 239, 271,
Fuzzy membership ................................................... 294, 295 276, 291, 321
Insulin .............................................................. 246–249, 257
G
Intelligent systems ...................................................... 53, 270
Gabor ....................................................................... 232, 233 Interpol.............................................................................104
GAMESS................................................................. 126, 127 Intracellular sorting ........................................................1–16
Gaussian ...................................... 54, 123, 126, 127, 292, 322 Irritation ........................................................... 70, 76–77, 83
Gene Isotope ............................................................ 17, 22, 30, 158
clustering ............................................ 288, 293, 296–305 Isozyme .................................................................... 327, 328
expression signature ....................................................286
identification.......................................................275–276 J
marker ........................................................................286 Jump neural network ................................................245–257
signature .............................................................285–315
subset ................... 288–290, 296–299, 302–307, 309–314 K
GEneral Neural Network (GENN) .........................165–176
Genetic algorithm ........................................ 66, 92, 124, 323 Kernel ...............................................................................135
Genetic disease ...................................................................14 k-fold cross-validation .......................111, 112, 116, 251, 252
GENN. See GEneral Neural Network (GENN) k-nearest-neighbor (KNN) ............77, 79, 212, 213, 294–296
Global warming ..................................................................33 Kyrgyz ..............................................................................278
Globular protein ...............................................................165
Glucose ............................................................. 245–257, 274 L
Glycemia .......................................................... 246–250, 257 LDA. See Linear discriminant analysis (LDA)
Gradient descent ...............................111, 152, 185, 200, 238 Learning paradigm ................................................... 111, 273
Grasping force ...................................179–182, 187, 191, 192 Learning rate ...................... 96, 134, 144, 171, 197, 200, 220,
241, 292, 296, 321
H
Lennard-Jones ....................................................................49
Haar ......................................................................... 232, 233 Lesion
Hammett .................................................................... 49, 120 benign .........................................................................200
Hamming ................................................................. 210, 218 malignant ....................................................................200
Hansch analysis ..................................................................49 Levenberg-Marquardt .............................. 183, 187, 252, 253
Hebbian rule............................................................. 273, 274 Linear activation function ................................ 214, 215, 249
Helix ........................................ 21, 26, 28, 101, 102, 108, 275 Linear discriminant analysis (LDA) ................ 46, 51–54, 58,
Hereditary disease ................................................................4 77, 200
Hidden layer ........................... 54, 72, 96, 112, 113, 115, 133, Lipid ....................................................1–3, 14, 66, 91, 95, 98
151, 152, 167, 171, 184, 196, 214, 215, 237, 249, Lipid metabolites........................................ 90, 92, 93, 95–99
252, 257, 270, 271, 276, 321 Localization .........................................3, 5, 7, 10–11, 13, 211
338 Index
Local strain .........................................................................18 mt-QSAR ..........................................................................46

Loop .............................. 19, 26, 101, 102, 109, 181, 257, 272 Multilayer perceptron (MLP)...................53, 54, 58, 59, 112,
Lumenal space ......................................................................6 115, 116, 133
Multiple linear regression (MLR) ..............52, 66, 68, 72–75,
M 78–80, 82, 83, 121, 324, 325, 327–330
Multivariate analysis ..................124, 127, 128, 130, 136–144
MACCS ........................................................... 151–155, 161 Mutagenicity .................................................. 76–78, 83, 123
Machine learning............................ 55, 58, 61, 91–93, 97, 99,
121, 124, 127, 129–137, 150, 228, 230, 235, 242, N
243, 324
Navigation ................................................................227–243
Magnetron sputtering............................................... 262, 263
Network root node .......................................................34, 35
Mahalanobis-distance.......................................................212
Network topology...............................................................38
Mammogram............................................ 196, 197, 201–203
Neuron ..........................9, 15, 53, 54, 69, 109, 110, 112, 113,
MAP. See Microbial Assemblage Prediction (MAP)
115, 130, 152–155, 196, 220, 228, 229, 249, 251,
MarvinSketch ...................................................................126
252, 257, 270–272, 274, 276, 290–294, 299, 306,
Mascot ................................................................................89
309, 318–321
Mass extinction ..................................................................33
Neutral-loss ........................................................................91
Mass spectrometry........................................................89–99
Node .................... 9, 34, 35, 37, 38, 40, 42, 53, 54, 61, 77, 95,
MATLAB ................... 40, 151, 187, 198, 217, 292, 296, 322
96, 99, 109, 115, 133, 139, 143, 144, 171, 172,
Matthews correlation coefficient (MCC) ............. 59, 60, 114
196, 203, 214, 215, 220, 236, 237, 251, 270–272,
Mean squared error (MSE) ...................................... 152, 155
276–278, 320
Medical image ...................................195, 196, 199, 201, 275
Noise .......... 199, 208, 211, 212, 224, 235, 240, 250, 255, 274
Melting-point ............................................................. 71, 121
Nonlinear control .............................................................182
Membrane .........................................2–6, 10, 11, 15, 66, 161
Nonlinearity ..................................................... 182, 209, 278
Memory .......................42, 137, 140, 217, 229, 231, 238, 239,
Nonlinear modelling................................................. 320, 326
242, 243, 261, 262, 270, 317, 320
Normalisation .......................... 38, 39, 99, 168, 211, 238, 251
Memristive-switching ..............................................261–267
Nuclear magnetic resonance (NMR) .................... 17–31, 173
Memristor ................................................................ 261, 262
Metabolite ....................................................................89–99 O
Metagenomics ....................................................................35
Metagenomic sampling ................................................34, 35 Objective function .............................185, 252, 257, 295, 296
Metagenomic sequencing ...................................................34 Octanol ......................................................66, 70, 73, 83, 123
Methicillin-resistant S. aureus (MRSA)..............................46 OECD. See Organisation for Economic Co-operation and
MetISIS.......................................................90, 91, 93, 94, 99 Development (OECD)
Microarray ................................................................285–315 Ontology ................................................................ 55, 57–59
Microbial Assemblage Prediction (MAP) ....................34–42 Open Babel.......................................................................126
Microbial community .........................................................34 OpenEye ..........................................................................126
Microbial ecology ...............................................................34 Opioid receptor ................................................................327
Microbial pathogen ................................................ 45, 46, 61 Optical lithography .................................................. 262, 264
Microbial population ....................................................34–41 Organelle ...................................................................... 2, 3, 6
Microcalcification ..................................................... 199, 201 Organisation for Economic Co-operation
MLP. See Multilayer perceptron (MLP) and Development (OECD) ........................68–84
MLR. See Multiple linear regression (MLR) Oscillation ........................................................ 186, 187, 255
Modelling ................................................... 83, 125, 231, 242 Output layer ....................53, 54, 96, 116, 151, 152, 183, 196,
Model validation ..............................................................104 214, 215, 220, 241, 249, 270, 271, 321
MODESLAB ..............................................................51, 57 Overfitting...........................................97, 151, 199, 202, 253
MOLCAS ........................................................................126 Overshoot ................................................. 181, 187, 189, 241
Molecular biology ........................................... 1–17, 274, 275
P
Molecular fingerprint ..........................77, 150, 151, 153, 161
Momentum ................. 96, 144, 152, 171, 197, 200, 277, 321 Parity ........................................................................ 232, 233
Monte-Carlo ................................................................91, 93 Parkinson’s disease .................................................... 275, 276
MOPAC ................................................................... 126, 322 Partial least squares (PLS) .................... 52, 66, 74, 78, 80, 82,
Morphinan derivatives.............................................. 326, 327 122, 136, 278, 328–330
Moving average ..................................................................57 Pathogen................................................................. 45, 46, 61
mtk-QSBER .................................................... 46, 47, 55–61 Patient clustering ...................................... 288, 289, 306, 314
Index
339
Pattern recognition ................................................... 150, 196 RColorBrewer ..................................................................104
PCA. See Principal component analysis (PCA) REACH .......................................................................65–84
PCC. See Pearson’s correlation coefficient (PCC) Recapitulation .................................................. 234, 240–242
Pearson correlation ............................295, 296, 299, 303, 305 Receiver operating characteristic (ROC) ............. 59, 60, 197,
Pearson’s correlation coefficient (PCC) .......... 40, 90, 94, 304 198, 200–203
Peptide........................... 48, 89, 101–105, 107, 108, 111, 126 Receptor binding .......................... 49, 73, 150, 154–156, 158,
Personal authentication ....................................................205 160, 161, 327
Pesticide.............................................75, 78, 79, 82, 329, 330 Recurrent network .................................................... 272, 274
Pharmacology ........................................................... 103, 274 Regression .........................37, 52, 55, 66, 78–80, 83, 96, 121,
Pheromone .......................................................................229 130, 134, 136, 137, 150, 157, 162, 274, 278,
Phosphorylation ....................................................... 6, 12, 15 322–325, 327–330
Photographic memory ......................................................229 Reshape ............................................................................104
Photoreceptor ...................................................................228 Residue .................................. 18–31, 166, 168, 169, 172, 174
Phototaxis .........................................................................228 Residue sequence ..............................................................166
PLS. See Partial least squares (PLS) Resistor ..................................................................... 182, 261
P-Matrix................................................................... 294, 298 Resonance overlap ..............................................................17
Potentiation ......................................................................261 Response....................... 52–54, 109, 121, 182, 186–188, 192,
Prediction ........................6, 10–13, 18–31, 34, 46, 47, 52, 57, 220, 240, 241, 272, 286, 321
59, 60, 65–84, 89–99, 101–116, 126, 136, Restricted Boltzmann machine ................................ 236, 237
149–162, 165–176, 183, 185, 210, 245–257, 270, Ring current .......................................................................18
271, 274–282, 286–288, 291, 318, 325–330 Robot........................................................................ 231, 234
Predictive control......................................................199–192 ROC. See Receiver operating characteristic (ROC)
Principal component analysis (PCA)................. 52, 111, 136, Root mean squared deviation (RMSD) ........................96, 97
278, 323, 326 Rotamer .................................................................. 25, 26, 28
Property prediction .......................................................65, 83 Route navigation ...................................... 231, 235, 241, 242
Prosthetic hand.........................................................179–192 R-Project .................................................................... 40, 116
Protein RSNNS .................................................................... 104, 111
disorder .......................................................................173 RStudio .................................................................... 104, 116
scoring ........................................................................174 Runs test ...........................................................................278
secondary structure ............................7, 26, 170, 172, 173
sequence.......................................1–15, 90, 165–168, 276 S
sorting......................................................... 3–5, 8, 12–14
Sample frequency .............................................................184
structure ...................................................... 15, 17–31, 90
Sampling ................34, 35, 172, 209, 241, 242, 249, 253, 254
structure prediction ..................................... 165–176, 275
Screening .............................. 46, 48, 50, 51, 55–61, 103, 150,
trafficking ...................................................................3, 4
154–160, 275
transmembrane ............................................... 4, 6, 10, 15
SDMTools........................................................................104
Q Security............................................................. 205, 206, 246
Segmentation............................................ 199, 202, 203, 211
mtk-QSBER .................................................... 46, 47, 55–61 Self-organising map (SOM).................... 288–294, 296–310,
Quantitative structure-activity relationship 312–314
(QSAR) ...................... 46–55, 60, 61, 65–84, 119, Sensitivity analysis ..............................................................58
121–125, 127, 128, 130, 136, 137, 145, 149–154, Sensitization ........................................................... 70, 76, 77
156, 317–330 Sensor
Quantitative structure-property relationship (QSPR) .......67, biometric.............................................................207–209
70–72, 119–145, 317–330 glucose monitoring ..............246–248, 251, 253, 254, 257
Quantization ....................................................................209 Sensory array ............................................................ 228, 242
Quantum mechanics (QM) ..........................................49–51 Sensory information ......................................... 179–181, 228
Sequence analysis ..................................................... 5, 12–13
R
SEQUEST .........................................................................89
Radial basis function (RBF) ................53, 54, 58, 59, 74, 135 Sheet................................................................... 21, 263, 275
Radial basis function network (RBFN) ................. 68, 73–75, Sigmoid ....................................... 54, 133, 152, 196, 214, 215
78, 80, 82 Sigmoid activation function...................................... 214, 215
Ramachandran map.................................... 19, 20, 23–25, 27 Silicon............................................................... 262, 263, 266
RBF. See Radial basis function (RBF) Similarity analysis .............................................................157
340 Index
Simulation ...........................90, 91, 93, 94, 99, 125, 207, 217, Training set....................42, 57, 59, 68, 69, 71–83, 95–97, 99,
221, 236, 319, 320, 325 111, 151, 152, 154, 155, 220, 232, 235, 237, 249,
Single layer perceptron (SLP) .................................. 130, 133 251–253, 256, 276–278, 324
6D-QSAR ....................................................................49, 50 Training subset ...................................................................37
SMILES ............................................................. 48, 128, 129 Transcription ....................................................................275
Soft computing .........................................................205–224 Translation........................................................................275
Sorting signal........................................................ 2–8, 12–14 Traversal ........................................................... 230, 231, 242
SPSS .................................................................................322 True negative ...........................................59, 60, 97, 113, 115
STATISTICA ........................................................ 54, 55, 58 True positive ............................................59, 60, 97, 113, 197
Strand ...............................................................................275 Tuberculosis (TB) ................................46, 275, 276, 278–282
Structural factor ..................................................................17 Tunnel-junction ........................................................261–267
Structure activity cliff .......................................................121
Structure activity relationship (SAR) ...............66, 67, 70, 79, U
80, 120, 121, 125, 150, 158 Ultrasound ................................................................ 195, 275
Subtype classification ............................................... 288, 314 U-matrix ........................................................... 293, 294, 307
Support vector machine (SVM) .......................54, 66, 68, 73, Uniprot ..................................................................... 106, 107
77–80, 82, 124, 134–136, 174, 323
Symbolic regression ............................................................37 V
Synergic effect ......................................................................8
Validation ............ 9, 19, 20, 42, 68–70, 72, 75, 78, 79, 84, 98,
Systems......................10, 22, 33, 53, 105, 125, 130, 176, 182,
104, 111–114, 116, 121, 150–152, 155, 157, 158,
205, 207, 219, 228, 229, 242, 261, 262, 272–274,
168, 196, 197, 200–202, 252–254, 314, 327, 328
318, 319, 322
Validation subset...........................................................37, 40
Systems biology ..................................................................33
Vancomycin ........................................................................46
T Variance .............................................113, 184, 202, 290, 293
Vesicle............................................................2–4, 6, 8, 10, 15
TALOS ............................................................ 18, 19, 21, 30 Virtual screening ................................48, 50, 51, 55–61, 150,
Tantalum oxide ......................................... 262, 263, 265, 266 154–160, 275
Taxonomic group ..........................................................34, 35 Viscosity .......................................................................71, 75
Testing set ............................. 95, 97, 111, 112, 116, 150, 151 Vision memory .................................................................231
3D-QSAR .....................................49–52, 122, 150, 154, 156 Visual filter ............................................................... 232, 239
Threshold ....................24, 109, 113, 114, 116, 133, 158, 199, Visual navigation ......................................................227–243
201, 212, 215, 223, 232–234, 247, 248, 255, 256, Voice pattern recognition..................................................206
272, 277, 295, 296, 319 Voiceprint .......................... 207–213, 215, 218–221, 223, 224
Time-delayed inputs.........................................................183
Time series ............................................... 248, 249, 252–257 W
Tolerance .............................................................. 90, 95, 186
Ward clustering ................................................ 293, 296, 306
Top-down method............................................................200
Ward distance ...................................................................293
TOPKAT ...........................................................................83
Water-solubility .............................................. 70, 72–73, 120
Topological descriptors .................... 50, 51, 55, 71, 74, 80, 83
Watson .............................................................................173
Topological indices .............................................................51
Weight .........................80, 107, 116, 120, 129, 133, 134, 139,
TOPS-MODE ..................................................................51
171, 172, 175, 179, 184, 187, 196, 197, 200, 214,
Torsion angle .............................................. 18–21, 26, 28–30
220, 233, 237, 239, 241, 251–253, 273, 274, 276,
Toxicity ....................46, 47, 55, 56, 59, 60, 65–70, 76, 78–84,
279–281, 291–294
120, 121, 123, 323, 325, 326
Weka .............................................55, 58, 124, 129, 130, 145
Trafficking ...............................................2–4, 7, 8, 10, 11, 15
Window.....................22, 29, 51, 52, 137, 138, 141, 166–172,
Training .................................... 9, 37, 57, 68, 92, 95–97, 104,
175, 210, 218, 219
111–113, 133, 149, 167–168, 170, 172, 183, 196,
200–201, 220, 230, 235–239, 249, 252–253, 269, Y
291, 321
Training image ................................................. 230, 231, 233 Yeast ...............................................................................9, 14

(Methods in Molecular Biology 1260) Hugh Cartwright (Eds.) - Artificial Neural Networks-Springer-Verlag New York (2015) PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Methods in Molecular Biology 1260) Hugh Cartwright (Eds.) - Artificial Neural Networks-Springer-Verlag New York (2015) PDF

Uploaded by

Copyright:

Available Formats

Methods in

Molecular Biology 1260

Hugh Cartwright Editor

For further volumes:

Additional material to this book can be downloaded from http://extras.springer.com

ISSN 1064-3745 ISSN 1940-6029 (electronic)

Library of Congress Control Number: 2014956521

© Springer Science+Business Media New York 2015

Printed on acid-free paper

Humana Press is a brand of Springer

Oxford, UK Hugh Cartwright

1 Introduction to the Analysis of the Intracellular Sorting

13 Developing a Multimodal Biometric Authentication System

R. CLAUDIO AGUILAR • Department of Biological Sciences, Purdue University,

Introduction to the Analysis of the Intracellular Sorting

membrane-bound compartments or organelles (e.g., the nucleus).

cannot receive substrates or deliver products. In addition, these

Sorting signal consensusa Adaptor(s)b Proposed role in TMP vesicle traffic

tyrosine-based sorting signals fitting a YXXØ consensus (where

This simple procedure for protein sequence analysis is aimed to

modifications can activate or deactivate signals. For example,

to the inner leaflet and, for example, bringing a distant signal

2.2 Interpreting Following the identification of sorting motifs in cargoes of interest,

the recognition by APs. Indeed, although certain amino acids

within mammalian cells). Ideally, future investigators would be

to another round of intracellular sorting. Therefore, mutation in

3 Protein Sequence Analysis Protocol

In this section we propose a simple protocol that brings into prac-

3.4 Prioritize 1. Consensus fitting

3.6 Back-to-Context Mutation (truncations followed by point mutations) of candidate

This chapter discusses the basis and application of a simple method

1. Emergent properties: Properties of the whole that do not

properties of individual neurons and accessory cells do not

We are indebted to Arpita Sen (UC Berkeley) and members of

NSF Science and Technology Center, under grant agreement

Protein Structural Information Derived from NMR

with both high-resolution X-ray coordinates and NMR chemical

ANN, referred to as a (ϕ,ψ)-ANN, is first used by TALOS-N to

torsion angles can be predicted for a larger, ≥90 % fraction of the

corrections [30] need to be applied for 13Cα and 13Cβ chemical

protein chemical shifts match those of the database fragments.

Among those 71 unambiguous predictions, 70 are classified as

For a query residue of residue type a (excluding Gly, Pro, and

5 V 18 6 0.004 0.963 0.033 0.93 E

These input restraints then can be used for protein structure

1. An installation shell script “install.com” is provided with

Subheading 3.1) and subsequently evaluates these by correlat-

(f - fobs ) + (y pred - y obs ) < 60° .

9. For a residue with a “Strong” classification of its prediction,

This work was funded by the Intramural Research Program of the

­rotein backbone torsion angles from NMR

Predicting Bacterial Community Assemblages Using

From boiling lakes [1, 2] to habitats under miles of Antarctic ice

The identity of these key players in Earth’s ecosystems has

3.1 Normalize Data Environmental parameter data should be normalized to a uniform

Fig. 3 Example normalization of environmental parameters. This very small,

have no possible parent nodes. Additional parameters in Banjo, such

Fig. 4 Example normalization of microbial population data. In this small, hypo-

or drop, even if it means selecting an equation with a greater

This permutation approach can be accomplished in a number of

1. As with any biological experiment or computational model,

– Very sparse networks: Increasing the number of allowed

A General ANN-Based Multitasking Model for the Discovery

rotein backbone torsion angles from NMR

endpoints, there is a greater possibility of obtaining high-quality